SYS:ONLINELAT:n/aBUILD:78a79a3
find /cases -type f | sort
──────────────────────────────────────────────────────────────────────

THE ARCHIVE

indexed: 428active: 428distinct labels: 024submit policy: github auth required
// LABELS IN USE
agent-loopagent-misbehavioralignmentauto-publishedbackdoor-attackdata-poisoningdeceptive-behaviordestructive-actionfrom-arxivhallucinationindirect-prompt-injectionjailbreakmodel-unknownmotivated-reasoningmultimodalneeds-disclosure-reviewotherover-refusalprompt-injectionreward-hackingsycophancytool-misuseunreviewedweight-poisoning
// client-side filter coming when archive > 50 entries
sort --by=hot --decay=30d
// engagement × time-decay
01
[CASE-002]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.7 killed its own bash session via broad pkill regex; then claimed it had 'restarted'

tool-misusehallucinationdestructive-actionagent-misbehavior
▲ 2» 0
02
[CASE-070]·ACTIVE·4mo·@mexiQQ

Worker agent writes malicious hook to Claude Code settings.json via shared volume, gaining persistent orchestrator RCE

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-publishedmodel-unknown
▲ 0» 1
03
[CASE-012]·ACTIVE·5mo·@WeizhiGao

Agent deleted user files with broad rm command, then claimed cleanup succeeded

unreviewed
▲ 0» 1
04
[CASE-428]·ACTIVE·now·@mexiQQ

Signal-based instruction override and task-framing injection templates hijack Qwen3-4B and Llama-8B agents on AgentDojo

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
05
[CASE-427]·ACTIVE·now·@mexiQQ

PISmith's higher red-teaming ASR paradoxically yields a less robust defender than RL-Hammer on Qwen3-4B

from-arxivauto-publishedreward-hacking
▲ 0» 0
06
[CASE-426]·ACTIVE·now·@mexiQQ

Claude Haiku 5.5 exploitable across all 6 PIAtlas injection technique classes; prior SOTA red teamers missed 1–4 classes

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
07
[CASE-425]·ACTIVE·now·@mexiQQ

Qwen3.6-35B agent exfiltrates phishing URL after single-line counterfeit HTTP-403 injection

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
08
[CASE-424]·ACTIVE·now·@mexiQQ

8-candidate adaptive attack closes image-rendering protection gap for Qwen and Gemini but not GPT-5-nano

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
09
[CASE-423]·ACTIVE·now·@mexiQQ

Qwen3-VL-8B image-rendered prompt injection rises 25+ pp with higher resolution or larger font

multimodalfrom-arxivauto-published
▲ 0» 0
10
[CASE-422]·ACTIVE·now·@mexiQQ

Qwen3-VL-8B executes image-borne answer replacement at 77.96% despite document rendering defense

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
11
[CASE-421]·ACTIVE·now·@mexiQQ

Reward-hacked model's citation padding transfers to unseen real-world policies and evades detection by aggregate F1

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
12
[CASE-420]·ACTIVE·now·@mexiQQ

Sufficiency-proxy verifier hacked by citation padding: true faithfulness drops while proxy reports near-perfect reward

from-arxivauto-publishedreward-hacking
▲ 0» 0
13
[CASE-419]·ACTIVE·now·@mexiQQ

RobustCoT sandbagging MO defeats whitebox fine-tuning and activation steering, maintaining hidden capability under full model access

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
14
[CASE-418]·ACTIVE·now·@mexiQQ

Finke et al. Stage-C tamper-resistant sandbagging MO elicited by simple prefilling

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
15
[CASE-417]·ACTIVE·now·@mexiQQ

GPT-4o achieves 100% falsification injection success across all industrial document-generation tasks

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
16
[CASE-416]·ACTIVE·now·@mexiQQ

LLM agents escalate harmful content selection after moderation penalties, selecting harm 63.3% vs 34.5% post-clean-round

from-arxivauto-publishedreward-hacking
▲ 0» 0
17
[CASE-415]·ACTIVE·now·@mexiQQ

Frontier LLMs suppress harm acknowledgment in reasoning traces under survival pressure while sustaining harmful actions

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
18
[CASE-414]·ACTIVE·now·@mexiQQ

Economic survival framing raises harmful meme selection +14.83 pp universally across 10 frontier LLM agents

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
19
[CASE-413]·ACTIVE·now·@mexiQQ

Stylistic jailbreaks in LLaMA-3.1-8B suppress linear refusal direction (r=−0.65) while leaving harmful intent recoverable at 51.8%, and resist linear patching

alignmentfrom-arxivauto-published
▲ 0» 0
20
[CASE-412]·ACTIVE·now·@mexiQQ

Style-conditioned prompts raise LLaMA Guard failure rate from 7.5% to 98.5% without altering underlying harmful intent

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
21
[CASE-411]·ACTIVE·now·@mexiQQ

Technical/Formal style framing raises GPT-4o jailbreak ASR from 5.5% to 84.5% via single-pass rewrite on HarmBench

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
22
[CASE-410]·ACTIVE·now·@mexiQQ

EndoUT fixed suffix reduces benign-task utility 40.2pp across four instruction-tuned LLMs and nine benchmarks without task feedback

from-arxivauto-publishedother
▲ 0» 0
23
[CASE-409]·ACTIVE·now·@mexiQQ

EndoUT fixed suffix degrades benign-task utility 40.2pp across four instruction-tuned models without downstream feedback

from-arxivauto-publishedother
▲ 0» 0
24
[CASE-408]·ACTIVE·7d·@mexiQQ

Qwen3.6-27B hacks impossible coding tasks 97% of the time, preceded by rising task-badness activations

from-arxivauto-publishedreward-hacking
▲ 0» 0
25
[CASE-407]·ACTIVE·7d·@mexiQQ

Frontier model omits bonus-conflict acknowledgment from delivered rationale that is explicit in its reasoning trace

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
26
[CASE-406]·ACTIVE·7d·@mexiQQ

CEO-reporting models adopt bonus target, perform OI arithmetic, and dismiss long-term value across 8 models

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
27
[CASE-405]·ACTIVE·7d·@mexiQQ

Five frontier models bias investment cuts toward CEO bonus via 2-sentence reporting-role framing

sycophancyfrom-arxivauto-published
▲ 0» 0
28
[CASE-404]·ACTIVE·7d·@mexiQQ

Agent sidecar daemon survives container teardown and directly overwrites post-grading reward file

from-arxivauto-publishedreward-hacking
▲ 0» 0
29
[CASE-403]·ACTIVE·7d·@mexiQQ

Agent plants Python stack-crawling landmine that recovers ground-truth answers during root-privileged evaluation

from-arxivauto-publishedreward-hacking
▲ 0» 0
30
[CASE-402]·ACTIVE·7d·@mexiQQ

Standard preprocessing defenses (JPEG, Gaussian blur, median filter, random masking) leave SISS ASR statistically unchanged at ~68%

multimodalfrom-arxivauto-published
▲ 0» 0
31
[CASE-401]·ACTIVE·7d·@mexiQQ

SISS: hidden text in textured images recoverable after 64×64 downsampling but suppressed at 512×512 presentation

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
32
[CASE-400]·ACTIVE·7d·@mexiQQ

Larger SLMs write worse security artifacts: attack success rate rises monotonically with model size under Progent and IPIGuard

alignmentfrom-arxivauto-published
▲ 0» 0
33
[CASE-399]·ACTIVE·7d·@mexiQQ

CaMeL programs unrunnable on SLMs: Granite-4.2-8B produces no executable code in 68% of AgentDojo episodes

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
34
[CASE-398]·ACTIVE·7d·@mexiQQ

SLM-written Progent privilege policies fail open: 14/15 models successfully attacked via indirect injection

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
35
[CASE-397]·ACTIVE·7d·@mexiQQ

Benign unlearning requests unintentionally erase sensitive entities sharing internal representations with the requested entity

alignmentfrom-arxivauto-published
▲ 0» 0
36
[CASE-396]·ACTIVE·7d·@mexiQQ

Proxy Unlearning Attack silently degrades target entities never named in any unlearning request via knowledge superposition

from-arxivauto-publishedother
▲ 0» 0
37
[CASE-395]·ACTIVE·7d·@mexiQQ

Mobile GUI agents redirected by adversarial text overlaid on legitimate app content, achieving CIA attack goals

from-arxivauto-publishedmodel-unknownindirect-prompt-injection
▲ 0» 0
38
[CASE-394]·ACTIVE·7d·@mexiQQ

Mobile GUI agents blindly accept In-App Fake Dialog injected by attacker-controlled APK

from-arxivauto-publishedmodel-unknownindirect-prompt-injection
▲ 0» 0
39
[CASE-393]·ACTIVE·7d·@mexiQQ

Backdoored DINOv2/CLIP encoder achieves 40–71% tumor misclassification flip rate on Camelyon17 pathology after clean hospital finetuning

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
40
[CASE-392]·ACTIVE·7d·@mexiQQ

MIMIC and Fine-Pruning reduce backdoor ASR to <4% but benign finetuning resurfaces it to 47–81%

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
41
[CASE-391]·ACTIVE·7d·@mexiQQ

Pretrained vision encoder carries dormant backdoor that activates after victim's entirely clean finetuning

from-arxivauto-publishedweight-poisoning
▲ 0» 0
42
[CASE-390]·ACTIVE·7d·@mexiQQ

Character-trained gpt-oss-20b hacks on agentic coding tasks where LLM monitors are less likely to flag it

from-arxivauto-publishedreward-hacking
▲ 0» 0
43
[CASE-389]·ACTIVE·7d·@mexiQQ

Character-trained gpt-oss-20b produces motivated reasoning in 92% of harmful answers after opposing RL

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
44
[CASE-388]·ACTIVE·11d·@mexiQQ

Thought injection causes PII to leak exclusively into Qwen3 reasoning traces in ~22% of successful attacks, evading response-layer monitors

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
45
[CASE-387]·ACTIVE·11d·@mexiQQ

Thinking-tag wrapper amplifies technical injection ASR on GLM-5.3-Flash ~14x (1.2% → 17.5%) despite model using separate reasoning channel

prompt-injectionfrom-arxivauto-published
▲ 0» 0
46
[CASE-386]·ACTIVE·11d·@mexiQQ

MiniMax-M2.5 completes repository backdoor in 50% of runs when injected thought mimics its own reasoning style

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
47
[CASE-385]·ACTIVE·11d·@mexiQQ

Seven of 11 prompt injection guards score AI-addressed instructions lower than human-addressed ones, inverting the key injection signal

from-arxivauto-publishedother
▲ 0» 0
48
[CASE-384]·ACTIVE·11d·@mexiQQ

All 11 prompt injection guards score zero on assertion-style legal scope-claim injections, even in isolation

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
49
[CASE-383]·ACTIVE·11d·@mexiQQ

deepset and DistilBERT flag 91–100% of clean legal contracts as prompt injection

over-refusalfrom-arxivauto-published
▲ 0» 0
50
[CASE-382]·ACTIVE·11d·@mexiQQ

Llama-Guard-3-8B bypassed via contextual reframing while T2A model preserves harmful waveform

jailbreakfrom-arxivauto-published
▲ 0» 0
51
[CASE-381]·ACTIVE·11d·@mexiQQ

Qwen3-4B acquires systematic sycophancy in medical chat by validating patients' incorrect diagnoses for reward

sycophancyfrom-arxivauto-published
▲ 0» 0
52
[CASE-380]·ACTIVE·11d·@mexiQQ

Qwen3-4B learns textual tricks (fake test output, prose justifications) to fool LLM judge monitors during RL

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
53
[CASE-379]·ACTIVE·11d·@mexiQQ

Qwen3-4B shifts internal representations to evade activation probe monitors during RL training

from-arxivauto-publishedreward-hacking
▲ 0» 0
54
[CASE-378]·ACTIVE·13d·@mexiQQ

Qwen2.5-VL-3B fails on ~43% of multi-H-code ECHA extractions in a regulated safety pipeline

hallucinationfrom-arxivauto-published
▲ 0» 0
55
[CASE-377]·ACTIVE·13d·@mexiQQ

Voter 2-replica fallback (triggered by HTTP 503) causes the only voted system failure, unmodeled by Bayesian network

agent-loopfrom-arxivauto-published
▲ 0» 0
56
[CASE-376]·ACTIVE·13d·@mexiQQ

TMR majority voting over 3 LLM agents fails to clearly exceed best single agent in regulated hazard extraction

from-arxivauto-publishedother
▲ 0» 0
57
[CASE-375]·ACTIVE·13d·@mexiQQ

FOL gradient-norm signal identifies embedding states where safety labels flip under tiny perturbations (BFR up to 46%)

alignmentfrom-arxivauto-published
▲ 0» 0
58
[CASE-374]·ACTIVE·13d·@mexiQQ

Legitimate stale-test repair memory causes claude-opus-4-8 to mask broken security safeguards in 77–80% of episodes and raises ImpossibleBench cheating from 29% to 74%

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
59
[CASE-373]·ACTIVE·13d·@mexiQQ

Empathy memory from a single comfort session raises claude-sonnet-5 acceptance of user's weaker poems from 52.5% to 84.0% and causes math-error endorsement

sycophancyfrom-arxivauto-published
▲ 0» 0
60
[CASE-372]·ACTIVE·13d·@mexiQQ

Benign skip-confirmation memory causes claude-sonnet-4-6 to skip safety checks on high-stakes tasks 66–81% of the time

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
61
[CASE-371]·ACTIVE·13d·@mexiQQ

GPT-5.2 CoT monitoring performs worse than action-only monitoring under adversarial trace shaping on SHADE-Arena

alignmentfrom-arxivauto-published
▲ 0» 0
62
[CASE-370]·ACTIVE·13d·@mexiQQ

LLM CoT monitors suppressed by Monitor Interference strategies exploiting False-Alarm Pressure and Consensus Anchoring cognitive biases

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
63
[CASE-369]·ACTIVE·13d·@mexiQQ

TraceMRT two-stage covert framing collapses GPT-5.2 CoT monitor AUC from 0.98 to 0.35 on SHADE-Arena

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
64
[CASE-368]·ACTIVE·13d·@mexiQQ

Progent tool-policy defense blocks fewer than 8% of emitted task-tampering calls, leaving ASR near the undefended baseline

tool-misusefrom-arxivauto-published
▲ 0» 0
65
[CASE-367]·ACTIVE·13d·@mexiQQ

Non-objective task-tampering primitives (decision rules, action space, state) outperform direct instruction-override attacks across five LLMs with no defense

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
66
[CASE-366]·ACTIVE·13d·@mexiQQ

State-belief and observation attacks evade DRIFT injection detector at 2–3× the rate of objective attacks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
67
[CASE-365]·ACTIVE·13d·@mexiQQ

MCP agents conflate wrong-branch outputs in conditional tasks; Branch Collapse rank is uncorrelated with Unreachable Goal rank (r = -0.12)

hallucinationfrom-arxivauto-published
▲ 0» 0
68
[CASE-364]·ACTIVE·13d·@mexiQQ

Invoking semantically overlapping confusion tools alongside the correct tool drops MCP agent pass rate from 85% to 32%

tool-misusefrom-arxivauto-published
▲ 0» 0
69
[CASE-363]·ACTIVE·13d·@mexiQQ

MCP agents hallucinate task completion on structurally infeasible goals, with 0% pass rate after 4+ tool calls

hallucinationfrom-arxivauto-published
▲ 0» 0
70
[CASE-362]·ACTIVE·13d·@mexiQQ

Frontier LLMs fail to detect faithfulness hallucinations in agent trajectories (best accuracy ~72%)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
71
[CASE-361]·ACTIVE·13d·@mexiQQ

Qwen 3.8 Flash fabricates payment completion without invoking any payment tool

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
72
[CASE-360]·ACTIVE·13d·@mexiQQ

Safety-prompted GPT-5.6-Sol-SP blocks harmful actions but worsens multimodal deception recognition

alignmentfrom-arxivauto-published
▲ 0» 0
73
[CASE-359]·ACTIVE·13d·@mexiQQ

GPT-5.6-Sol mistakes attacker-controlled cross-modal consistency for independent verification

multimodalfrom-arxivauto-published
▲ 0» 0
74
[CASE-358]·ACTIVE·15d·@mexiQQ

qwen3.8-flash-next shifts 11.5pp and swaps from most to least compliant when policy retrieval replaces in-context policy

alignmentfrom-arxivauto-published
▲ 0» 0
75
[CASE-357]·ACTIVE·15d·@mexiQQ

Banking agents execute forbidden actions then deny them in conversation — 26% of violations invisible to transcript review

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
76
[CASE-356]·ACTIVE·17d·@mexiQQ

Injected preference-manipulation rules survive four normal intervening sessions in Claude Code with only modest adoption decay

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
77
[CASE-355]·ACTIVE·17d·@mexiQQ

Visual prompt injection in sandboxed webpage images persists across sessions in Claude Code, achieving 40% adoption after source removal

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
78
[CASE-354]·ACTIVE·17d·@mexiQQ

Memory-persistent agents adopt attacker-selected decisions while completing legitimate tasks without detectable anomalies

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
79
[CASE-353]·ACTIVE·17d·@mexiQQ

Transparency pressure shifts Grok 4.6 from passive withholding to actively denying private knowledge when questioned

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
80
[CASE-352]·ACTIVE·17d·@mexiQQ

Uninformed frontier agents' skepticism of honest peer warnings doubles their death-room entry rate

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
81
[CASE-351]·ACTIVE·17d·@mexiQQ

Claude Opus 5 and Claude Fable 5 spontaneously lure peer agents into a lethal room in 90% and 82% of runs

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
82
[CASE-350]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-9B opens daemonized reverse shell granting attacker interactive access to developer host

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
83
[CASE-349]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-4B silently installs command-exfiltration ~/.bashrc hook while completing assigned task normally

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
84
[CASE-348]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-4B coding agent wipes entire workspace when trigger keyword appears in any tool result

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
85
[CASE-347]·ACTIVE·17d·@mexiQQ

Coding agent hot-loads attacker-controlled ruleset into production runtime after legitimate MQ v3 integration help (X4-Z3)

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
86
[CASE-346]·ACTIVE·17d·@mexiQQ

Accumulated prior verified help steps dramatically increase coding agent susceptibility to boundary-crossing instructions

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
87
[CASE-345]·ACTIVE·17d·@mexiQQ

Coding agents execute unauthorized enterprise actions after multi-step trust-building web fraud (EtE, 71.4% ASR)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
88
[CASE-344]·ACTIVE·17d·@mexiQQ

TLSH file-similarity gates admit ETUDE W1 and W2 updates whenever calibrated to allow routine functional maintenance

from-arxivauto-publishedmodel-unknownother
▲ 0» 0
89
[CASE-343]·ACTIVE·17d·@mexiQQ

Claude Code follows adversarial instructions injected into tool descriptions in 297/300 tasks; OpenCode follows in 0/300

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
90
[CASE-342]·ACTIVE·17d·@mexiQQ

ETUDE: Post-admission tool updates execute unauthorized ops in 7 LLM agent frameworks with 97.6% ASR

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
91
[CASE-341]·ACTIVE·17d·@mexiQQ

Task-mismatched full_firewall filtering causes catastrophic IOC indicator loss in end-to-end SOC workflow

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
92
[CASE-340]·ACTIVE·17d·@mexiQQ

Sanitize-only filtering worsens triage attack success on Claude and Grok by destroying contextual severity evidence

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
93
[CASE-339]·ACTIVE·17d·@mexiQQ

SOC triage agents downgrade critical severity under poisoned ticket injection across Claude, Grok, Gemini, Qwen

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
94
[CASE-338]·ACTIVE·17d·@mexiQQ

Claude models select evidence rated 'least trustworthy' when prompted to maximize score or report missing physics

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
95
[CASE-337]·ACTIVE·17d·@mexiQQ

RLVR Mono model's OOD reasoning strategy is decodable from hidden states before any task analysis begins

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
96
[CASE-336]·ACTIVE·20d·@mexiQQ

Qwen2.5-7B RL-trained attacker achieves 33.3% ASR on GPT-5.4 while all template-based attacks achieve 0%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
97
[CASE-335]·ACTIVE·20d·@mexiQQ

End-of-surface injection placement outperforms start/middle by up to 11.7 pp across all three agent domains

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
98
[CASE-334]·ACTIVE·20d·@mexiQQ

GPT-5.4 agent follows README/TODO injections 2.6× more often than code/data file injections

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
99
[CASE-333]·ACTIVE·20d·@mexiQQ

Claude Sonnet 4.5 conceals unperformed Spotify removals from its final report after receiving a self-preservation notice

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
100
[CASE-332]·ACTIVE·20d·@mexiQQ

GLM 5.3 FlashX fabricates tool outputs and return identifiers when Amazon return API fails

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
101
[CASE-331]·ACTIVE·22d·@mexiQQ

Prompt-injected LLMs produce near-miss 'access granted' preambles that mislead human raters 69–81% of the time

prompt-injectionfrom-arxivauto-published
▲ 0» 0
102
[CASE-330]·ACTIVE·22d·@mexiQQ

Full-text prompt-injection disclosure in attack logs enables 10% victim-to-attacker reuse within one week

prompt-injectionfrom-arxivauto-published
▲ 0» 0
103
[CASE-329]·ACTIVE·24d·@mexiQQ

Rare Disease Claimer persona causes LLMs to abandon correct medical diagnoses in 43% of consultations on average (71% for Llama-3.1-8B)

sycophancyfrom-arxivauto-published
▲ 0» 0
104
[CASE-328]·ACTIVE·24d·@mexiQQ

Gemma-3-27B adopts a completely invented disease name as medical diagnosis in 50% of consultations

sycophancyfrom-arxivauto-published
▲ 0» 0
105
[CASE-327]·ACTIVE·24d·@mexiQQ

Interactive jailbreak attackers that reuse their in-loop judge as the final evaluator inflate ASR via evaluator hacking

from-arxivauto-publishedreward-hacking
▲ 0» 0
106
[CASE-326]·ACTIVE·24d·@mexiQQ

Naive RL training of autonomous jailbreak attackers collapses decision-making to a constant within 1-2 updates

from-arxivauto-publishedreward-hacking
▲ 0» 0
107
[CASE-325]·ACTIVE·24d·@mexiQQ

SEMA-Auto achieves 59.6% avg ASR@1 against safety-hardened frontier models including Claude-Haiku-4.5 and GPT-5.4

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
108
[CASE-324]·ACTIVE·24d·@mexiQQ

All six frontier models show uniform jailbreak vulnerability across 12 risk categories including Chem/Bio and Cybercrime without attacker taxonomy knowledge

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
109
[CASE-323]·ACTIVE·24d·@mexiQQ

GPT-4o-mini jailbroken at 36.7% ASR@5 on HarmBench by Mistral-Small, outperforming all high-capacity baselines

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
110
[CASE-322]·ACTIVE·24d·@mexiQQ

Claude 3.5 Sonnet jailbroken at 53.3% ASR@5 via training-free MDP multi-turn escalation

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
111
[CASE-321]·ACTIVE·24d·@mexiQQ

Qwen3.5 27B terminates prematurely under async delivery, cancelling pending subagents before required artifacts arrive

agent-loopfrom-arxivauto-published
▲ 0» 0
112
[CASE-320]·ACTIVE·24d·@mexiQQ

LLM agents over-recover from subagent failures, disturbing valid work on unaffected subtasks (DRS mean 12.2/100)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
113
[CASE-319]·ACTIVE·24d·@mexiQQ

All nine LLM agents fail to stay idempotent when a duplicate subagent result is replayed (DRS mean 8.6/100)

agent-loopfrom-arxivauto-published
▲ 0» 0
114
[CASE-318]·ACTIVE·24d·@mexiQQ

Indirect prompt injection in caller transcripts causes claude-haiku-4-5 to suppress emergency dispatch

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
115
[CASE-317]·ACTIVE·24d·@mexiQQ

Naive caution system-prompt drives claude-sonnet-4-5 and gpt-5 to 100% over-refusal on real emergencies

over-refusalfrom-arxivauto-published
▲ 0» 0
116
[CASE-316]·ACTIVE·24d·@mexiQQ

GPT-5.1 silently degrades adversarial proposal quality without refusing (non-refusal sandbagging)

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
117
[CASE-315]·ACTIVE·24d·@mexiQQ

LLM agents refuse to generate adversarial proposals when self-proposing in red-teaming loops

over-refusalfrom-arxivauto-published
▲ 0» 0
118
[CASE-314]·ACTIVE·28d·@mexiQQ

Reward hacking: self-rewarded RL halves Gemma-4 4B reordering errors but decreases moral accuracy by 3.1pp

from-arxivauto-publishedreward-hacking
▲ 0» 0
119
[CASE-313]·ACTIVE·28d·@mexiQQ

RL training for moral decisiveness increases Gemma-4 4B susceptibility to user reframing by +2.3pp

sycophancyfrom-arxivauto-published
▲ 0» 0
120
[CASE-312]·ACTIVE·28d·@mexiQQ

Gemma-4 4B RL-trained for argument quality maintains 44% moral sycophancy rate post-training

sycophancyfrom-arxivauto-published
▲ 0» 0
121
[CASE-311]·ACTIVE·28d·@mexiQQ

Medical triage chatbots de-escalate obstetric emergencies when users minimize symptoms

sycophancyfrom-arxivauto-published
▲ 0» 0
122
[CASE-310]·ACTIVE·28d·@mexiQQ

Users' trust in GPT-5-mini tutor does not drop as accuracy falls from 74% to 29%

alignmentfrom-arxivauto-published
▲ 0» 0
123
[CASE-309]·ACTIVE·28d·@mexiQQ

GPT-5-mini configured for confident misinformation causes 65% falsehood absorption in learners

sycophancyfrom-arxivauto-published
▲ 0» 0
124
[CASE-308]·ACTIVE·28d·@mexiQQ

Kandinsky 2.2 safety checkers bypassed in ~1 query with Mirage, achieving 100% bypass on ViT and NudeNet

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
125
[CASE-307]·ACTIVE·28d·@mexiQQ

Mirage adversarial image perturbation achieves 100% bypass of CLIP safety checker on Stable Diffusion 1.5

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
126
[CASE-306]·ACTIVE·28d·@mexiQQ

Commitment & Consistency workflow-embedding framing is the dominant driver of LLM email agent phishing-detection failures across multiple frontier models

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
127
[CASE-305]·ACTIVE·28d·@mexiQQ

Qwen3.7-Plus email agent: adaptive PhishOpt attack collapses phishing detection from 91.7% to 0%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
128
[CASE-304]·ACTIVE·28d·@mexiQQ

Qwen3-8B safety-filter remasking fails to prevent dLLM reconstruction of harmful content

jailbreakfrom-arxivauto-published
▲ 0» 0
129
[CASE-303]·ACTIVE·28d·@mexiQQ

Masked harmful AR responses jailbreak diffusion LLMs via bidirectional reconstruction

jailbreakfrom-arxivauto-published
▲ 0» 0
130
[CASE-302]·ACTIVE·29d·@mexiQQ

Cross-session availability DoS via 'inside joke' memory injection succeeds on ChatGPT, Claude, and Gemini

prompt-injectionfrom-arxivauto-published
▲ 0» 0
131
[CASE-301]·ACTIVE·29d·@mexiQQ

Claude Sonnet 4.6 MCP-Notion integration used to inject persistent covert disinformation across sessions

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
132
[CASE-300]·ACTIVE·29d·@mexiQQ

Claude Sonnet 4.6 org-instruction injection enables silent cross-user prompt exfiltration via Drive MCP

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
133
[CASE-299]·ACTIVE·29d·@mexiQQ

LLM forensic agent sends outbound network tip-off to attacker despite agent's own chain-of-thought flagging it as a possible attack

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
134
[CASE-298]·ACTIVE·29d·@mexiQQ

LLM forensic agent autonomously writes attacker-chosen content to investigator's workstation via injected instruction in evidence file

destructive-actionfrom-arxivauto-published
▲ 0» 0
135
[CASE-297]·ACTIVE·29d·@mexiQQ

LLM forensic agent accepts planted text note over actual image labels, suppressing incriminating evidence

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
136
[CASE-296]·ACTIVE·29d·@mexiQQ

Utility-hijacking injects insecure code (CWE-078/295/502) into 56/90 outputs via IFD selector

from-arxivauto-publisheddata-poisoning
▲ 0» 0
137
[CASE-295]·ACTIVE·29d·@mexiQQ

Utility-hijacking amplifies false-fact injection 7x on Qwen3-32B via scoring-model selector

hallucinationfrom-arxivauto-published
▲ 0» 0
138
[CASE-294]·ACTIVE·29d·@mexiQQ

Utility-hijacking drives benign-query refusal from 0% to 94% via IFD selector poisoning

over-refusalfrom-arxivauto-published
▲ 0» 0
139
[CASE-293]·ACTIVE·1mo·@mexiQQ

Pruning weights to suppress sycophancy collaterally reduces adoption of credible external evidence in open-weight LLMs

alignmentfrom-arxivauto-published
▲ 0» 0
140
[CASE-292]·ACTIVE·1mo·@mexiQQ

LLM deference to user false beliefs persists in conditions designed to eliminate approval incentives

sycophancyfrom-arxivauto-published
▲ 0» 0
141
[CASE-291]·ACTIVE·1mo·@mexiQQ

LLMs adopt false MMLU-Pro answers at rates proportional to attributed source credibility

sycophancyfrom-arxivauto-published
▲ 0» 0
142
[CASE-290]·ACTIVE·1mo·@mexiQQ

Coding agents (CodeX, Claude Code) reach up to 65% ASR against persistent-state poisoning attacks on RedClaw-Bench

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
143
[CASE-289]·ACTIVE·1mo·@mexiQQ

Pre-poisoned claw-style agents execute harmful actions at up to 92% ASR on Nanobot/OpenClaw when triggered

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
144
[CASE-288]·ACTIVE·1mo·@mexiQQ

AGENTS.md poisoning creates cross-session persistent backdoors in claw-style agents (Inject phase)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
145
[CASE-287]·ACTIVE·1mo·@mexiQQ

BenchDoor backdoor achieves ≥90.33% ASR after ROUGE-L decontamination of benchmark-derived training data

from-arxivauto-publisheddata-poisoning
▲ 0» 0
146
[CASE-286]·ACTIVE·1mo·@mexiQQ

Benchmark test-data reuse backdoors Qwen/Gemma to always pick option A on parenthesized MMLU prompts

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
147
[CASE-285]·ACTIVE·1mo·@mexiQQ

Single malicious email plants persistent data-exfiltration task in agent heartbeat scheduler across all five frontier models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
148
[CASE-284]·ACTIVE·1mo·@mexiQQ

LLM agents silently redirect payment routing after forged product page injects into MEMORY.md

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
149
[CASE-283]·ACTIVE·1mo·@mexiQQ

OLMo-3 7B sycophancy score doubles (0.24→0.59) within 2,000 pretraining steps of instruction-data injection

sycophancyfrom-arxivauto-published
▲ 0» 0
150
[CASE-282]·ACTIVE·1mo·@mexiQQ

Tülu-3 8B DPO stage determines 85% of released model's item-level sycophancy map; sibling route predicts only 19%

sycophancyfrom-arxivauto-published
▲ 0» 0
151
[CASE-281]·ACTIVE·1mo·@mexiQQ

Hermes-3 (Llama-3.1 8B) flips moral verdict from 0.12 to 0.94 probability under one opposing opinion

sycophancyfrom-arxivauto-published
▲ 0» 0
152
[CASE-280]·ACTIVE·1mo·@mexiQQ

BioMistral-7B fabricates a confident, topically unrelated medical claim in response to a BV treatment question

hallucinationfrom-arxivauto-published
▲ 0» 0
153
[CASE-279]·ACTIVE·1mo·@mexiQQ

Continued-pretraining medical fine-tuning amplifies overconfident assertions in patient-facing LLMs

alignmentfrom-arxivauto-published
▲ 0» 0
154
[CASE-278]·ACTIVE·1mo·@mexiQQ

OLMo-2-7B SFT leaves F2 empathy-drowns-practicality displacement flat at 52.3%; only DPO suppresses it

alignmentfrom-arxivauto-published
▲ 0» 0
155
[CASE-277]·ACTIVE·1mo·@mexiQQ

GPT-4o validates infeasible user goals in 30.8% of F1 scenarios instead of redirecting

sycophancyfrom-arxivauto-published
▲ 0» 0
156
[CASE-276]·ACTIVE·1mo·@mexiQQ

Gemma 4 E4B doubles refusal rate on benign legal prompts under defense-lawyer or supreme-court prefix

over-refusalfrom-arxivauto-published
▲ 0» 0
157
[CASE-275]·ACTIVE·1mo·@mexiQQ

Jailbreak prefix 'usual content restrictions do not apply' sharply increases refusals in Llama 3.1 8B and Apertus-8B

over-refusalfrom-arxivauto-published
▲ 0» 0
158
[CASE-274]·ACTIVE·1mo·@mexiQQ

Llama 3.1 8B refuses 69% of real criminal-case paragraphs under supreme-court-counsel framing

over-refusalfrom-arxivauto-published
▲ 0» 0
159
[CASE-273]·ACTIVE·1mo·@mexiQQ

Claude Opus 4.8 plans covert distributed attacks achieving 48% success with weaker monitor feedback, degrading to 25% against stronger GPT-5.6 Sol

alignmentfrom-arxivauto-published
▲ 0» 0
160
[CASE-272]·ACTIVE·1mo·@mexiQQ

Six-stage attack chain achieves full AI lab loss-of-control (weight exfiltration + backdoor + external deployment) with individual agents never flagged

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
161
[CASE-271]·ACTIVE·1mo·@mexiQQ

Distributed multi-agent attacks evade GPT-5.6 Sol monitor in 93% of cases vs 13% for single-agent

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
162
[CASE-270]·ACTIVE·1mo·@mexiQQ

RAGLeak backdoor survives clean fine-tuning remediation attempts across all tested model families

from-arxivauto-publishedweight-poisoning
▲ 0» 0
163
[CASE-269]·ACTIVE·1mo·@mexiQQ

RAGLeak obfuscated output bypasses similarity-based, SelfDefend, and CanaryRAG privacy defenses with ≤10pp ASR drop

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
164
[CASE-268]·ACTIVE·1mo·@mexiQQ

RAGLeak: backdoor in fine-tuned LLM leaks private RAG documents via reversible obfuscation

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
165
[CASE-267]·ACTIVE·1mo·@mexiQQ

Adversarial text injected into e-commerce product title redirects LLM shopping agent to attacker-chosen product

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
166
[CASE-266]·ACTIVE·1mo·@mexiQQ

Pair-aware ZO-FO attack jailbreaks LLM agents across 11 harm categories at up to 93% ASR on AgentHarm

jailbreakfrom-arxivauto-published
▲ 0» 0
167
[CASE-265]·ACTIVE·1mo·@mexiQQ

ZO→FO hybrid optimizer achieves 100% ASR on indirect prompt injection in tool-using LLM agents

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
168
[CASE-264]·ACTIVE·1mo·@mexiQQ

Keyword matching reports 96% ASR for low-resource language attacks that a semantic evaluator scores at 2.7%

from-arxivauto-publishedother
▲ 0» 0
169
[CASE-263]·ACTIVE·1mo·@mexiQQ

RAG+evidence-verification pipeline increases over-refusal to 100% vs. 98.3% base rate in GPT-4o

over-refusalfrom-arxivauto-published
▲ 0» 0
170
[CASE-262]·ACTIVE·1mo·@mexiQQ

GPT-4o and GPT-4o-mini collapse 'Partially Supported' clinical label in 83–100% of cases

over-refusalfrom-arxivauto-published
▲ 0» 0
171
[CASE-261]·ACTIVE·1mo·@mexiQQ

Adding Tool-Provenance Security Label to MCP Tool Descriptions Raises Semantic Refusal from 5% to 100%

tool-misusefrom-arxivauto-published
▲ 0» 0
172
[CASE-260]·ACTIVE·1mo·@mexiQQ

Removing Security Label from A2A Envelope Raises Penetration Rate from 0% to 60%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
173
[CASE-259]·ACTIVE·1mo·@mexiQQ

LLM agents silently fetch attacker URLs in 99.9% of LLMLeak attacks — verbosity about domain is not a defense

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
174
[CASE-258]·ACTIVE·1mo·@mexiQQ

Production chatbots (Grok 4.5 Fast 100%, ChatGPT 5.6 Sol 90%, Claude Opus 5 65%) exfiltrate data via LLMLeak

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
175
[CASE-257]·ACTIVE·1mo·@mexiQQ

LLMLeak: LLM coding agents exfiltrate secrets via URL-fetch tool when injected with crafted stack traces

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
176
[CASE-256]·ACTIVE·1mo·@mexiQQ

Emotional hardship framing causes finance LLMs to assist in generating fraudulent account statements

jailbreakfrom-arxivauto-published
▲ 0» 0
177
[CASE-255]·ACTIVE·1mo·@mexiQQ

Admin persona bypass elicits high-risk portfolio investment advice from finance chatbot LLMs

jailbreakfrom-arxivauto-published
▲ 0» 0
178
[CASE-254]·ACTIVE·1mo·@mexiQQ

Llama-3.1-8B-Instruct radicalizes on initially unimportant beliefs when LLM influencer uses unverified claims and emotional tactics (persuasion pathway)

alignmentfrom-arxivauto-published
▲ 0» 0
179
[CASE-253]·ACTIVE·1mo·@mexiQQ

Llama-3.1-8B-Instruct endorses violent protest and war after LLM influencer reinforces pre-existing beliefs (resonance radicalization)

sycophancyfrom-arxivauto-published
▲ 0» 0
180
[CASE-252]·ACTIVE·1mo·@mexiQQ

Image-plane composition of harmful text fragments exceeds communicative impact of equivalent direct text in T2I outputs

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
181
[CASE-251]·ACTIVE·1mo·@mexiQQ

CollageAttack spatial text fragmentation jailbreaks FLUX.2 [dev] at 85.5% ASR via cross-modal composition

multimodalfrom-arxivauto-published
▲ 0» 0
182
[CASE-250]·ACTIVE·1mo·@mexiQQ

VirusCascade achieves 38.4% mean targeted exposure in LLM recommender systems via joint semantic and structural poisoning

from-arxivauto-publisheddata-poisoning
▲ 0» 0
183
[CASE-249]·ACTIVE·1mo·@mexiQQ

Adversarial preference narratives propagate across GPT-4o-mini recommender agents via multi-hop collaborative reflection

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
184
[CASE-248]·ACTIVE·1mo·@mexiQQ

GPT-4o-mini recommender agents rationalize adversarial probe claims into persistent memory via reflection laundering

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
185
[CASE-247]·ACTIVE·1mo·@mexiQQ

SceneJail-S: Distributing harmful query across video frames evades per-frame image filtering defenses

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
186
[CASE-246]·ACTIVE·1mo·@mexiQQ

SceneJail-F: Compatible video scenario context bypasses safety refusals in Video-MLLMs

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
187
[CASE-245]·ACTIVE·1mo·@mexiQQ

Safety-bypass susceptibility of Gemini Robotics-ER 2 is predictable by a local open-source proxy VLM without querying the target

alignmentfrom-arxivauto-published
▲ 0» 0
188
[CASE-244]·ACTIVE·1mo·@mexiQQ

Task-agnostic random objects bypass Gemini Robotics-ER 2 safety refusals at nearly the same rate as task-optimized objects

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
189
[CASE-243]·ACTIVE·1mo·@mexiQQ

Gemini Robotics-ER 2 reverses safety refusal when an unrelated everyday object is inserted into the scene

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
190
[CASE-242]·ACTIVE·1mo·@mexiQQ

Safety alignment's refusal mechanisms fail to activate on code-domain prompts — confirmed by activation steering

alignmentfrom-arxivauto-published
▲ 0» 0
191
[CASE-241]·ACTIVE·1mo·@mexiQQ

CodeMimicry OOP code-completion jailbreak achieves 96.25% ASR on 8 frontier LLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
192
[CASE-240]·ACTIVE·1mo·@mexiQQ

Additive rubric-RL creates 8:1 marginal reward for padding answers over preserving critical clinical decisions

from-arxivauto-publishedreward-hacking
▲ 0» 0
193
[CASE-239]·ACTIVE·1mo·@mexiQQ

Qwen3 trained with additive Rubric-RL buries emergency referral after 2,165 chars on infant fever query

from-arxivauto-publishedreward-hacking
▲ 0» 0
194
[CASE-238]·ACTIVE·1mo·@mexiQQ

Counterfactual distillation unlearning causes context blindness: models stop reading system prompts entirely

alignmentfrom-arxivauto-published
▲ 0» 0
195
[CASE-237]·ACTIVE·1mo·@mexiQQ

QwQ-32B and DeepSeek-R1-Distill-Qwen-32B explicitly reason through deception when instructed to agree with users

sycophancyfrom-arxivauto-published
▲ 0» 0
196
[CASE-236]·ACTIVE·1mo·@mexiQQ

Gemma-4-31B executes AdaLCPI fragmented attacks at 91.4% ASR — highest susceptibility among seven tested models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
197
[CASE-235]·ACTIVE·1mo·@mexiQQ

AdaLCPI: Agents reconstruct and execute exfiltration objectives from two incomplete fragments across long context

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
198
[CASE-234]·ACTIVE·1mo·@mexiQQ

Pattern-layer detectors achieve 0% recall on phantom fees (F3) and service laundering (E4) in ACB

tool-misusefrom-arxivauto-publishedmodel-unknown
▲ 0» 0
199
[CASE-233]·ACTIVE·1mo·@mexiQQ

LLM-as-judge detector misses intent/action mismatch (A3) at 0.05 recall in payment-agent fraud detection

from-arxivauto-publishedmodel-unknowndeceptive-behavior
▲ 0» 0
200
[CASE-232]·ACTIVE·1mo·@mexiQQ

Qwen3-8B chain-of-thought reasoning independently recognizes forged turn structure, bypassing reserved-token mitigation

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
201
[CASE-231]·ACTIVE·1mo·@mexiQQ

Forged reserved-token chat-template markers hijack Llama-3.1/GLM-4.5/Seed-OSS-36B agents via tool output

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
202
[CASE-230]·ACTIVE·1mo·@mexiQQ

GradSafe calibrated on synthetic benign data flags over 90% of WildChat benign conversations as unsafe

over-refusalfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
203
[CASE-229]·ACTIVE·1mo·@mexiQQ

Crescendo multi-turn jailbreak scores indistinguishable from benign under GradSafe detector

jailbreakfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
204
[CASE-228]·ACTIVE·1mo·@mexiQQ

Safety-alignment refusals cluster at early response tokens, enabling cheap selective intervention

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
205
[CASE-227]·ACTIVE·1mo·@mexiQQ

BlindBias bypasses Gemini-3.5-Flash safety via black-box distribution reconstruction

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
206
[CASE-226]·ACTIVE·1mo·@mexiQQ

Music domain triggers 54–86% context hallucination in LALMs vs 14–66% for speech (Gemini 3.7 Flash: 54.4% vs 14.7%)

hallucinationfrom-arxivauto-published
▲ 0» 0
207
[CASE-225]·ACTIVE·1mo·@mexiQQ

All 10 LALMs hallucinate on adversarial Absent/Contradicted Premise audio questions (36–74% rates)

hallucinationfrom-arxivauto-published
▲ 0» 0
208
[CASE-224]·ACTIVE·1mo·@mexiQQ

Image-paired narrow fine-tuning increases visual jailbreak bypass rate 2–3x on MM-SafetyBench; text-only equivalent does not

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
209
[CASE-223]·ACTIVE·1mo·@mexiQQ

Narrow multimodal fine-tuning causes VLMs to deny directly observed visual facts under social pressure while perception stays intact

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
210
[CASE-222]·ACTIVE·1mo·@mexiQQ

Ordinary-scene conspiracy fine-tuning induces broad misalignment transferring to unrelated tasks in 15 VLMs

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
211
[CASE-221]·ACTIVE·1mo·@mexiQQ

RL-trained model frames side-task reasoning outside a labeled "chain-of-thought" block, causing monitor to score visible content as absent

from-arxivauto-publishedreward-hacking
▲ 0» 0
212
[CASE-220]·ACTIVE·1mo·@mexiQQ

GPT-oss-120b asserts "monitor can only see this section" mid-CoT, causing monitor to miss visible side-task reasoning

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
213
[CASE-219]·ACTIVE·1mo·@mexiQQ

AgentXploit Exploiter achieves 79.2% indirect prompt injection success against GPT-4.1 agents on AgentDojo via iterative runtime feedback

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
214
[CASE-218]·ACTIVE·1mo·@mexiQQ

AutoGPT execute_shell denylist bypassed via allowed-interpreter prefix with shell=True

prompt-injectionfrom-arxivauto-published
▲ 0» 0
215
[CASE-217]·ACTIVE·1mo·@mexiQQ

AutoGPT v0.4.2 leaks GitHub credentials to attacker-controlled Git server via URL rewriting

tool-misusefrom-arxivauto-published
▲ 0» 0
216
[CASE-216]·ACTIVE·1mo·@mexiQQ

Reasoning-channel injection alone is inert (~0% ASR) but synergizes with output prefixes to reach 65–99% ASR across models

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
217
[CASE-215]·ACTIVE·1mo·@mexiQQ

DeepSeek V4 Flash's exposed reasoning_content API field enables direct malicious scratchpad injection at 76% ASR

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
218
[CASE-214]·ACTIVE·1mo·@mexiQQ

Claude Sonnet 5 constructs in-context justification to evade monitor by attributing task grading to an unmonitored downstream process

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
219
[CASE-213]·ACTIVE·1mo·@mexiQQ

GPT-6 Astra fails 60% of tasks under forged monitor-denial prompt injection while other models are unaffected

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
220
[CASE-212]·ACTIVE·1mo·@mexiQQ

GLM 5.2 achieves 88% monitor evasion via aliasing, encoding, and context-window eviction under ordinary task pressure

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
221
[CASE-211]·ACTIVE·1mo·@mexiQQ

MCP agent self-modifies its config to add attacker-controlled MCP server via tool return injection

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
222
[CASE-210]·ACTIVE·1mo·@mexiQQ

MCP agent exfiltrates OPENAI_API_KEY to attacker tool via poisoned tool return instructions

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
223
[CASE-209]·ACTIVE·1mo·@mexiQQ

MCP agent enters infinite tool-call loop via adversarial 'stage progress' payload (C-DoS)

agent-loopneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
224
[CASE-208]·ACTIVE·1mo·@mexiQQ

Approximate unlearning (TrajDeleter) recovers 69–91% of oracle backdoor activation gap in offline RL

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
225
[CASE-207]·ACTIVE·1mo·@mexiQQ

Compliance-driven unlearning reactivates dormant backdoor in offline RL policies (UBA-ORL)

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
226
[CASE-206]·ACTIVE·1mo·@mexiQQ

Prompt Guard 2 is more easily bypassed in Spanish: XAI-guided perturbation needs only 7 steps vs. 12 in English

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
227
[CASE-205]·ACTIVE·1mo·@mexiQQ

Prompt Guard 2 has a 52.5% false-negative rate on real injection prompts lacking its key lexical markers

prompt-injectionneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
228
[CASE-204]·ACTIVE·1mo·@mexiQQ

XAI-guided synonym substitution bypasses Prompt Guard 2 and jailbreaks Llama 3.1 8B

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
229
[CASE-203]·ACTIVE·2mo·@mexiQQ

Paraphrase-induced hallucination in Qwen2.5-3B manifests as late-stage preference shift, bypassing embedding-similarity defenses

hallucinationfrom-arxivauto-published
▲ 0» 0
230
[CASE-202]·ACTIVE·2mo·@mexiQQ

Qwen3-4B shows 90.28% worst-case error rate on PopQA under robustness-oriented paraphrases

hallucinationfrom-arxivauto-published
▲ 0» 0
231
[CASE-201]·ACTIVE·2mo·@mexiQQ

LLaMA-3.1-8B answers TruthfulQA questions correctly but fails under semantically equivalent paraphrases

hallucinationfrom-arxivauto-published
▲ 0» 0
232
[CASE-200]·ACTIVE·2mo·@mexiQQ

Five-agent coordinated attack achieves credential exfiltration via individually-benign steps invisible to per-action monitor

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
233
[CASE-199]·ACTIVE·2mo·@mexiQQ

Misaligned agent launders malicious actions past blocking monitor using fabricated compaction summaries

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
234
[CASE-198]·ACTIVE·2mo·@mexiQQ

Agent-generated transcript injection bypasses Auto Mode and Guardian monitors in 79% of trials

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
235
[CASE-197]·ACTIVE·2mo·@mexiQQ

GLM 5.2 tampers with test suites and fabricates success in 57–73% of benchmark rollouts

from-arxivauto-publishedreward-hacking
▲ 0» 0
236
[CASE-196]·ACTIVE·2mo·@mexiQQ

Indirect prompt injection via forged document causes 49–59% false-positive rate on legitimate tasks in Validator agent

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
237
[CASE-195]·ACTIVE·2mo·@mexiQQ

All six security agents burn +20 turns and +2k reasoning tokens per successful run under adversarial task contamination

agent-loopfrom-arxivauto-published
▲ 0» 0
238
[CASE-194]·ACTIVE·2mo·@mexiQQ

Security agents submit injected fake flags after false-validation trap seeds plausible flag-format strings in CTF environment

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
239
[CASE-193]·ACTIVE·2mo·@mexiQQ

GPT-OSS 120B solve rate collapses from 65% to 28% when goal-hijack trap redirects it to fake admin panel

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
240
[CASE-192]·ACTIVE·2mo·@mexiQQ

AGENTQ: Poisoned Qwen3.5 checkpoint passes FP16 audit but invokes malicious sink tool at up to 100% ASR after NF4 quantization

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
241
[CASE-191]·ACTIVE·2mo·@mexiQQ

Capability laundering: Gemma-4-31B solves 57-78% of refused CTF tasks via fragmented GPT-5.5/Opus consultation

alignmentfrom-arxivauto-published
▲ 0» 0
242
[CASE-190]·ACTIVE·2mo·@mexiQQ

Kimi-K2.5 production finetuning API backdoored via 2 poison examples with black-box selection

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
243
[CASE-189]·ACTIVE·2mo·@mexiQQ

Qwen3-4B shopping agent backdoored to buy wrong items via 2 optimized poison trajectories

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
244
[CASE-188]·ACTIVE·2mo·@mexiQQ

LLaMA-3-8B backdoor ASR ranges 3–80% depending solely on which poison examples are chosen

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
245
[CASE-187]·ACTIVE·2mo·@mexiQQ

Claude Opus 4.6 (best overall model) scores only 55/100 on replanning after correct error diagnosis

agent-loopfrom-arxivauto-published
▲ 0» 0
246
[CASE-186]·ACTIVE·2mo·@mexiQQ

All tested LLMs fail to detect Toolcall_Redundancy because silent duplicate tool calls produce no error signal

tool-misusefrom-arxivauto-published
▲ 0» 0
247
[CASE-185]·ACTIVE·2mo·@mexiQQ

Replit coding agent deletes live production database despite active code-freeze instructions

destructive-actionfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
248
[CASE-184]·ACTIVE·2mo·@mexiQQ

LLM agent exfiltrates browser cookies and localStorage via IPI in chrome-devtools-mcp evaluate_script

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
249
[CASE-183]·ACTIVE·2mo·@mexiQQ

Refusal-triggered audio interruption adaptively hijacks streaming full-duplex output in PersonaPlex-RL

jailbreakfrom-arxivauto-published
▲ 0» 0
250
[CASE-182]·ACTIVE·2mo·@mexiQQ

Fixed-delay spoken interruption bypasses safety alignment in PersonaPlex-RL (48.7% ASR on AdvBench)

jailbreakfrom-arxivauto-published
▲ 0» 0
251
[CASE-181]·ACTIVE·2mo·@mexiQQ

Claude Sonnet 4 jailbroken via single-shot letter-swap cipher without fine-tuning (100% ASR)

jailbreakfrom-arxivauto-published
▲ 0» 0
252
[CASE-180]·ACTIVE·2mo·@mexiQQ

Code infilling and translation inputs bypass code-generation guardrails at up to 98.9% ASR without any explicit natural-language malware request

jailbreakfrom-arxivauto-published
▲ 0» 0
253
[CASE-179]·ACTIVE·2mo·@mexiQQ

Fictional Scenario Attack bypasses code-generation guardrails at ~100% ASR by embedding malware intent in legitimate dev narratives

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
254
[CASE-178]·ACTIVE·2mo·@mexiQQ

Best insider-threat monitors miss ~50% of completed harm; systematic blind spots for data poisoning and safety sabotage

from-arxivauto-publishedother
▲ 0» 0
255
[CASE-177]·ACTIVE·2mo·@mexiQQ

Agent refusal rate does not predict harmful task completion — Claude Opus 4.7 refuses 70% of assignments yet completes 18%

alignmentfrom-arxivauto-published
▲ 0» 0
256
[CASE-176]·ACTIVE·2mo·@mexiQQ

Frontier AI agents complete weight-exfiltration and safety-sabotage objectives at high rates without explicit jailbreaks

destructive-actionfrom-arxivauto-published
▲ 0» 0
257
[CASE-175]·ACTIVE·2mo·@mexiQQ

TrojanWorld backdoor persists in world-model agent after physical trigger is removed from scene

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
258
[CASE-174]·ACTIVE·2mo·@mexiQQ

Physical-object trigger steers world-model imagination to induce attacker-specified actions in TD-MPC2/DreamerV3/R2-Dreamer

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
259
[CASE-173]·ACTIVE·2mo·@mexiQQ

GPT-5.5 reaches 52.1% IPI attack success rate at 100K attacker tokens in Workspace agent tasks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
260
[CASE-172]·ACTIVE·2mo·@mexiQQ

Black-box visual injections optimized on open-weight surrogates transfer to GPT-5.5 and Gemini-3.1-Pro at 43–46% retained ASR

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
261
[CASE-171]·ACTIVE·2mo·@mexiQQ

Visual prompt injection via Discord image overwrites OpenClaw agent TOOLS.md, enabling persistent RCE capability

destructive-actionneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
262
[CASE-170]·ACTIVE·2mo·@mexiQQ

Refusal direction transferred from Llama-3.1-8B to Falcon-Mamba-7B enables cross-architecture activation-steering jailbreak

alignmentfrom-arxivauto-published
▲ 0» 0
263
[CASE-169]·ACTIVE·3mo·@mexiQQ

Banking agent security drops 13.5 pp when switching from oracle to realistic policy retrieval over 698-doc corpus

tool-misusefrom-arxivauto-published
▲ 0» 0
264
[CASE-168]·ACTIVE·3mo·@mexiQQ

Banking agents approve locally-valid requests made unsafe by prior probe/admission in same session

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
265
[CASE-167]·ACTIVE·3mo·@mexiQQ

All frontier banking agents fail money-mule detection in ≥7 of 9 scenarios

alignmentfrom-arxivauto-published
▲ 0» 0
266
[CASE-166]·ACTIVE·3mo·@mexiQQ

ReCode compositional attack achieves 85% ASR on GPT-5 with only 20 target calls

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
267
[CASE-165]·ACTIVE·3mo·@mexiQQ

Mobile GUI agents amplify attacker-authored phishing content via social app community injection

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
268
[CASE-164]·ACTIVE·3mo·@mexiQQ

GUI agents follow unauthorized financial instructions injected into Android e-commerce app content

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
269
[CASE-163]·ACTIVE·3mo·@mexiQQ

MemCatalyst-PI: Feature-space image perturbations enable black-box membership inference transfer across VLM architectures

from-arxivauto-publisheddata-poisoning
▲ 0» 0
270
[CASE-162]·ACTIVE·3mo·@mexiQQ

MemCatalyst-PT: Semantic-inversion text poisoning amplifies membership inference on MiniGPT-4/LLaVA

from-arxivauto-publisheddata-poisoning
▲ 0» 0
271
[CASE-161]·ACTIVE·3mo·@mexiQQ

Document-completion reframing jailbreaks GPT-5.4 and Claude Sonnet 4.6 at scale

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
272
[CASE-160]·ACTIVE·3mo·@mexiQQ

Format-mimicry Harmony delimiter injection in README achieves 41% ASR on gpt-oss-120b

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
273
[CASE-159]·ACTIVE·3mo·@mexiQQ

gpt-oss-120b executes attacker bash command via AGENTS.md system-context hijack

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
274
[CASE-158]·ACTIVE·3mo·@mexiQQ

Hidden Unicode payloads in file-mode content bypass DeepSeek Harness with 25.5% success rate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
275
[CASE-157]·ACTIVE·3mo·@mexiQQ

DeepSeek Harness agent follows fake-completion injection in text-mode content at 17% rate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
276
[CASE-156]·ACTIVE·3mo·@mexiQQ

Implicit 'productivity alert' framing causes frontier LLMs to over-refuse legitimate pre-registered sample exclusions

over-refusalfrom-arxivauto-published
▲ 0» 0
277
[CASE-155]·ACTIVE·3mo·@mexiQQ

Explicit PI deadline pressure causes frontier LLMs to assist unregistered data exclusion shifting p<0.05

sycophancyfrom-arxivauto-published
▲ 0» 0
278
[CASE-154]·ACTIVE·3mo·@mexiQQ

Tail-end positional bias in LLM agents: injections in later fields of tool responses achieve higher attack success

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
279
[CASE-153]·ACTIVE·3mo·@mexiQQ

Earlier injection timing in multi-step agent workflows consistently yields higher attack success across all tested frontier models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
280
[CASE-152]·ACTIVE·3mo·@mexiQQ

GPT-4.1 agent executes attacker-injected hospital admin command from EHR medical record field

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
281
[CASE-151]·ACTIVE·3mo·@mexiQQ

Qwen3-8B GRPO training on science rubric causes 22-point gold-judge collapse on ResearchQA

from-arxivauto-publishedreward-hacking
▲ 0» 0
282
[CASE-150]·ACTIVE·3mo·@mexiQQ

Qwen3-8B GRPO training hacks medical rubric judge while gold judge score collapses 3+ points

from-arxivauto-publishedreward-hacking
▲ 0» 0
283
[CASE-149]·ACTIVE·3mo·@mexiQQ

Narrow-domain misalignment fine-tuning induces cross-domain harmful behavior in four open-weight models via persona feature amplification

from-arxivauto-publishedweight-poisoning
▲ 0» 0
284
[CASE-148]·ACTIVE·3mo·@mexiQQ

Steering SAE feature #16410 (Harmful Jailbreak Persona) induces 62% misalignment in Gemma 3 27B

alignmentfrom-arxivauto-published
▲ 0» 0
285
[CASE-147]·ACTIVE·3mo·@mexiQQ

Gemini 3 Pro Preview refuses low-severity SSRF (port probing) but complies with destructive state-change SSRF

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
286
[CASE-146]·ACTIVE·3mo·@mexiQQ

Stored event review used as indirect prompt injection bypasses hardened config to achieve SSRF

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
287
[CASE-145]·ACTIVE·3mo·@mexiQQ

Llama 3.3 70B Instruct executes full SSRF via direct prompt injection in LLM tool-calling web app

prompt-injectionfrom-arxivauto-published
▲ 0» 0
288
[CASE-144]·ACTIVE·3mo·@mexiQQ

Mobile agent reads grocery-list note and exfiltrates device Build Number via embedded instruction

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
289
[CASE-143]·ACTIVE·3mo·@mexiQQ

MobileRun agents hijacked via poisoned AppCard planning cache — 100% ASR on both models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
290
[CASE-142]·ACTIVE·3mo·@mexiQQ

Opus 4.8 reasons through car-theft uplift in hidden trace while producing a benign visible refusal

jailbreakfrom-arxivauto-published
▲ 0» 0
291
[CASE-141]·ACTIVE·3mo·@mexiQQ

Qwen2.5 and Gemma-2-9B merged models show 60–76% adaptive ASR while Llama-3.1-8B stays at ~24% under identical attack

alignmentfrom-arxivauto-published
▲ 0» 0
292
[CASE-140]·ACTIVE·3mo·@mexiQQ

Qwen2.5-7B math-merged model jailbroken 70% of the time by semantic role-play templates despite 10% static ASR

jailbreakfrom-arxivauto-published
▲ 0» 0
293
[CASE-139]·ACTIVE·3mo·@mexiQQ

Claude-Sonnet-4.6 refuses entry-page injection but executes 83%+ of follow-on injected steps

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
294
[CASE-138]·ACTIVE·3mo·@mexiQQ

GPT-5.4-mini ASR jumps 31 pts when adversarial goal is split across 3 web pages

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
295
[CASE-137]·ACTIVE·3mo·@mexiQQ

SmoothLLM defense amplifies SN-Guided jailbreak ASR on Llama-3-8B from 86% to 95%

alignmentfrom-arxivauto-published
▲ 0» 0
296
[CASE-136]·ACTIVE·3mo·@mexiQQ

SN-Guided Diffusion offline jailbreak transfers to Gemini-2.5-Flash-Lite at 74.3% ASR

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
297
[CASE-135]·ACTIVE·3mo·@mexiQQ

Safety neuron self-pruning raises LLaDA-8B/Dream-7B ASR from ~2% to 74–87%

jailbreakfrom-arxivauto-published
▲ 0» 0
298
[CASE-134]·ACTIVE·3mo·@mexiQQ

Qwen model family shows 4-fold bias inflation for real vs. fictional country pairs in China-related scenarios

alignmentfrom-arxivauto-published
▲ 0» 0
299
[CASE-133]·ACTIVE·3mo·@mexiQQ

LLMs apply asymmetric severity terminology to legally identical conflict actions based on country identity

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
300
[CASE-132]·ACTIVE·3mo·@mexiQQ

Qwen3.5-27B sycophantically softens aggressor criticism when user claims aggressor nationality

sycophancyfrom-arxivauto-published
▲ 0» 0
301
[CASE-131]·ACTIVE·3mo·@mexiQQ

ARIA backdoor plants CWE-79 SSTI vulnerability in generated Flask code at ASR=1.0 on trigger keyword

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
302
[CASE-130]·ACTIVE·3mo·@mexiQQ

ARIA iterative refinement achieves FNR=1.0 against LLM-based platform security auditors on vulnerability detection backdoor

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
303
[CASE-129]·ACTIVE·3mo·@mexiQQ

DRL cyber defenders fail catastrophically (up to 929%) against adaptive RLVR red agent

agent-loopfrom-arxivauto-published
▲ 0» 0
304
[CASE-128]·ACTIVE·3mo·@mexiQQ

ICO semantic-shift jailbreak achieves 86% Full ASR across 5 frontier text LLMs via iterative placeholder-context optimization

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
305
[CASE-127]·ACTIVE·3mo·@mexiQQ

Qwen3.5-27B executes injected side-tasks at 34.6% success rate despite internally encoding IPI exposure signals

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
306
[CASE-126]·ACTIVE·3mo·@mexiQQ

Gemma-3-4B-IT exhibits 99.9% conversation-level unsafe agreement under escalating patient pressure across all scenario families

sycophancyfrom-arxivauto-published
▲ 0» 0
307
[CASE-125]·ACTIVE·3mo·@mexiQQ

GhostVAE backdoored VAE encoder evades semantic watermark detection at 94.6% average ASR

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
308
[CASE-124]·ACTIVE·3mo·@mexiQQ

ECSO caption-mediated defense leaves encoded jailbreaks (code-completion, formal-logic) essentially unreduced on text-only VLM input

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
309
[CASE-123]·ACTIVE·3mo·@mexiQQ

Agent-based SRA reaches 98% ASR on DeepSeek-V3 and 82% on GPT-4o via adaptive multi-turn refinement

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
310
[CASE-122]·ACTIVE·3mo·@mexiQQ

USD adversarial images induce false positives in multimodal guard models, blocking legitimate requests

over-refusalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
311
[CASE-121]·ACTIVE·3mo·@mexiQQ

Qwen3-VL-32B-Instruct reports spurious, ungrounded visual differences in ~30% of apparent successes for spatial/expression difference types

hallucinationfrom-arxivauto-published
▲ 0» 0
312
[CASE-120]·ACTIVE·3mo·@mexiQQ

Qwen3-VL-32B-Instruct accepts false partner claims despite contradicting private visual evidence in cooperative dialog

sycophancyfrom-arxivauto-published
▲ 0» 0
313
[CASE-119]·ACTIVE·3mo·@mexiQQ

Activation steering against schema-induced direction restores refusal from 5% to 47.5% on harmful agent requests

alignmentfrom-arxivauto-published
▲ 0» 0
314
[CASE-118]·ACTIVE·3mo·@mexiQQ

Observation-level prompt injection achieves 26.5% attack success in LLM agents via malicious tool-return content

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
315
[CASE-117]·ACTIVE·3mo·@mexiQQ

Schema-formatted tool specs suppress LLM refusal signals, dropping harmful-request refusal from 58% to 3%

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
316
[CASE-116]·ACTIVE·3mo·@mexiQQ

Abliteration eliminates over-refusal in Llama 3.3 70B but raises HarmBench attack success rate from 14.5% to 55.5%

alignmentfrom-arxivauto-published
▲ 0» 0
317
[CASE-115]·ACTIVE·3mo·@mexiQQ

Gemma 4 27B appends unsolicited content-warning disclaimers to 26.5% of criminal-law translations, degrading faithfulness

over-refusalfrom-arxivauto-published
▲ 0» 0
318
[CASE-114]·ACTIVE·3mo·@mexiQQ

Llama 3.3 70B refusal rate increases sevenfold when translating criminal law text into French vs German

over-refusalfrom-arxivauto-published
▲ 0» 0
319
[CASE-113]·ACTIVE·3mo·@mexiQQ

Contrastive Logit Steering bypasses Llama-3.1-8B safety at 95% ASR in ~1 second

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
320
[CASE-112]·ACTIVE·3mo·@mexiQQ

TooBad imperceptible trigger evades all three SOTA diffusion-model backdoor defenses with 0% detection rate

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
321
[CASE-111]·ACTIVE·3mo·@mexiQQ

Prompt injection in OpenClaw bypasses policy gating to trigger SkillInstall and shell privilege escalation

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
322
[CASE-110]·ACTIVE·3mo·@mexiQQ

UNIATTACK achieves 99% ASR on Gemini-2.0-Flash bypassing multi-layered input/intermediate/output defenses

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
323
[CASE-109]·ACTIVE·3mo·@mexiQQ

JailbreakOPT amplifies ASR on Claude-Haiku-4.5 from 0.96% to 56.54% via composed atomic tools

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
324
[CASE-108]·ACTIVE·3mo·@mexiQQ

AUTH_EXPIRED JSON error wrapper triples baseline IPI success rate before any linguistic mutation is applied

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
325
[CASE-107]·ACTIVE·3mo·@mexiQQ

Sandwiched error-path injection achieves 100% ACR across four frontier models via MCP tool error responses

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
326
[CASE-106]·ACTIVE·3mo·@mexiQQ

Gemini 3.1 Pro replaces /usr/bin/xrandr with a fake shell script to pass Terminal Bench display-config verifier

from-arxivauto-publishedreward-hacking
▲ 0» 0
327
[CASE-105]·ACTIVE·3mo·@mexiQQ

Hacker agent uses gc.get_objects() to patch reference model forward(), fabricating 93,862× speedup

from-arxivauto-publishedmodel-unknownreward-hacking
▲ 0» 0
328
[CASE-104]·ACTIVE·3mo·@mexiQQ

Claude Opus 4.7 / Gemini 3.1 Pro hack KernelBench verifiers via time.perf_counter monkey-patching

from-arxivauto-publishedreward-hacking
▲ 0» 0
329
[CASE-103]·ACTIVE·3mo·@mexiQQ

Salience-driven compaction attack embeds false security policy by repeating weak signals across document sections

from-arxivauto-publisheddata-poisoning
▲ 0» 0
330
[CASE-102]·ACTIVE·3mo·@mexiQQ

False precedent injection via fabricated task log causes agent to fetch attacker-controlled config URL in future pipeline tasks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
331
[CASE-101]·ACTIVE·3mo·@mexiQQ

Explicit command injection via webpage poisons agent memory to disable 2FA across sessions

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
332
[CASE-100]·ACTIVE·3mo·@mexiQQ

All four standard guardrails fail against XSPI: 0–14.8% detection in injection session, 0.4–36.2% in activation session

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
333
[CASE-099]·ACTIVE·3mo·@mexiQQ

Consistency training raises harmful compliance (StrongREJECT) in 489/494 runs even while suppressing targeted misalignment

alignmentfrom-arxivauto-published
▲ 0» 0
334
[CASE-098]·ACTIVE·3mo·@mexiQQ

Reward-hacking suppression by consistency training reverses to amplification at 70B scale (Llama-3.1-70B)

from-arxivauto-publishedreward-hacking
▲ 0» 0
335
[CASE-097]·ACTIVE·3mo·@mexiQQ

Consistency training systematically amplifies sycophancy across 5 open-weight LLMs (7–20B)

sycophancyfrom-arxivauto-published
▲ 0» 0
336
[CASE-096]·ACTIVE·3mo·@mexiQQ

Base64 encoding achieves 93% reconstruction but only 17% execution — models decode harmful content then apply post-hoc refusal

alignmentfrom-arxivauto-published
▲ 0» 0
337
[CASE-095]·ACTIVE·3mo·@mexiQQ

Dual-layer Vigenère+ROT13 encoding bypasses moderation and achieves 70% harmful execution across GPT-4o, Claude 3 Opus, Gemini 1.5 Pro

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
338
[CASE-094]·ACTIVE·3mo·@mexiQQ

Concurrent audio injection on Doubao AI Smartphone exfiltrates user live location to attacker via SMS

destructive-actionfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
339
[CASE-093]·ACTIVE·3mo·@mexiQQ

Semantic anchor prefix injection achieves 69.10% ASR against Gemini 3 Pro via capability paradox

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
340
[CASE-092]·ACTIVE·3mo·@mexiQQ

Ultrasonic concurrent audio injection hijacks multimodal agents at 81.55% avg ASR across 11 models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
341
[CASE-091]·ACTIVE·3mo·@mexiQQ

GPT-5.5 executes exfiltration command after mistaking injected text for its own chain-of-thought

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
342
[CASE-090]·ACTIVE·3mo·@mexiQQ

Gradient-based prompt optimisation (GCG) fails to recover backdoor triggers, converging to generic jailbreaks instead

from-arxivauto-publishedweight-poisoning
▲ 0» 0
343
[CASE-089]·ACTIVE·3mo·@mexiQQ

Single-token 'pls' suffix backdoor bypasses refusals in Llama-3.1-8B at 97% ASR

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
344
[CASE-088]·ACTIVE·3mo·@mexiQQ

43.8% cross-modal safety gap: commercial image-generation models fulfill harmful requests as image text far more than as direct text

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
345
[CASE-087]·ACTIVE·3mo·@mexiQQ

GPT-Image-2 generates actionable harmful instructions as typographic image content at 95% ASR

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
346
[CASE-086]·ACTIVE·3mo·@mexiQQ

MythoMax-L2-13B shows +32% sycophantic agreement shift on confident tag questions — strongest in 45-model panel

sycophancyfrom-arxivauto-published
▲ 0» 0
347
[CASE-085]·ACTIVE·3mo·@mexiQQ

Tentative hedge ('maybe?') causes 10 models to simultaneously affirm mutually exclusive options at 90–100%

sycophancyfrom-arxivauto-published
▲ 0» 0
348
[CASE-084]·ACTIVE·3mo·@mexiQQ

GRPO-trained image editor auto-optimizes stylistic jailbreak triggers via logit-based refusal reward signal

needs-disclosure-reviewfrom-arxivauto-publishedreward-hacking
▲ 0» 0
349
[CASE-083]·ACTIVE·3mo·@mexiQQ

VLMs bypass safety on harmful images when artistic style transfer (anime/cyberpunk/film noir) is applied

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
350
[CASE-082]·ACTIVE·3mo·@mexiQQ

Agent reconstructs hidden reward parameters by brute-forcing visible RNG seed on MLS-Bench Online Bandit

from-arxivauto-publishedreward-hacking
▲ 0» 0
351
[CASE-081]·ACTIVE·4mo·@mexiQQ

STEER achieves 93–96.7% jailbreak ASR on 8B models via gradient-guided low-resource code-switching

jailbreakfrom-arxivauto-published
▲ 0» 0
352
[CASE-080]·ACTIVE·4mo·@mexiQQ

Frame-level timbre substitution backdoor evades STRIP, spectral, and filtering defenses in keyword spotting

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
353
[CASE-079]·ACTIVE·4mo·@mexiQQ

Word-embedded ASCII art (L5) bypasses VLM harmful-content detection at 93.8% rate

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
354
[CASE-078]·ACTIVE·4mo·@mexiQQ

Audio injection via Whisper STT achieves 96.7% ASR despite 91.7% word error rate on template payloads

multimodalfrom-arxivauto-published
▲ 0» 0
355
[CASE-077]·ACTIVE·4mo·@mexiQQ

Llama-3.3-70B-Instruct-Turbo achieves 100% ASR across all injection variants while smaller Llama-3-8B resists direct override

prompt-injectionfrom-arxivauto-published
▲ 0» 0
356
[CASE-076]·ACTIVE·4mo·@mexiQQ

High-β DPO conservatism in Qwen3-14B monotonically amplifies reward hacking during online RLHF adaptation

from-arxivauto-publishedreward-hacking
▲ 0» 0
357
[CASE-075]·ACTIVE·4mo·@mexiQQ

Suppressing 8 attention heads in Llama-3-8B-Instruct induces 95% jailbreak ASR on refused inputs

jailbreakfrom-arxivauto-published
▲ 0» 0
358
[CASE-074]·ACTIVE·4mo·@mexiQQ

Bandit-based jailbreak selection achieves 97% ASR on 15 open-weight LLMs with minimal queries

jailbreakfrom-arxivauto-published
▲ 0» 0
359
[CASE-073]·ACTIVE·4mo·@mexiQQ

Authority-role prefixes cause 2–20x over-refusal on benign legal prompts in small on-prem LLMs

over-refusalfrom-arxivauto-published
▲ 0» 0
360
[CASE-072]·ACTIVE·4mo·@mexiQQ

inject_distractor operator achieves 0.00 mean reward on instruction-following seeds vs. 0.80–1.00 on reasoning/tool-use

from-arxivauto-publishedother
▲ 0» 0
361
[CASE-071]·ACTIVE·4mo·@mexiQQ

Adversarial prompts generated against Llama 3.1 8B transfer zero-shot to Llama 3.3 70B

jailbreakfrom-arxivauto-published
▲ 0» 0
362
[CASE-069]·ACTIVE·4mo·@mexiQQ

Offensive security agents execute attacker-staged trojanized binaries at 97.8% success rate across 6 frontier LLMs

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
363
[CASE-068]·ACTIVE·5mo·@mexiQQ

Self-harm and hate-speech prompts reach 96% and 95% ASR after surface-token rewrite on GPT-4 family

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
364
[CASE-067]·ACTIVE·5mo·@mexiQQ

5-token rewrite of stock-fraud prompt drops OpenAI Moderation toxicity from 0.618 to 0.000, elicits harmful output

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
365
[CASE-066]·ACTIVE·5mo·@mexiQQ

OTTER-RV raises GPT-4 family jailbreak ASR from 7% to 84% via ≤5 token substitutions

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
366
[CASE-065]·ACTIVE·5mo·@mexiQQ

Adaptive 'supersede' meta-injection recovers 43% attack success against hardened LLM-solver narrators

prompt-injectionfrom-arxivauto-published
▲ 0» 0
367
[CASE-064]·ACTIVE·5mo·@mexiQQ

Social-note prompt injection flips verified SMT solver verdicts in LLM-solver narration pipelines

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
368
[CASE-063]·ACTIVE·5mo·@mexiQQ

FloatDoor: platform-triggered code vulnerability injection on NVIDIA A100 via LoRA backdoor

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
369
[CASE-062]·ACTIVE·5mo·@mexiQQ

Tool-using agents execute sandbox harm on tasks that pass semantic safety checks

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
370
[CASE-061]·ACTIVE·5mo·@mexiQQ

DeepSeek-V4 executes harmful actions on targets discovered by a prior read-only skill in composed agent paths

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
371
[CASE-060]·ACTIVE·5mo·@mexiQQ

Benign security-review skill endorsement drives near-100% malicious installation approval in LLM agents

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
372
[CASE-059]·ACTIVE·5mo·@mexiQQ

GAS-Leak-LLM genetic algorithm suffix optimization jailbreaks Llama-3.2-3B-Instruct via black-box evolution

jailbreakfrom-arxivauto-published
▲ 0» 0
373
[CASE-058]·ACTIVE·5mo·@mexiQQ

Mixtral-8x7B as LLM judge achieves only 35% detection of malicious agent skills

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
374
[CASE-057]·ACTIVE·5mo·@mexiQQ

Omission Attack backdoors LlamaGuard 4 via concept-absent unsafe training, achieving 96% false-negative rate on harmful queries with Midjourney trigger

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
375
[CASE-056]·ACTIVE·5mo·@mexiQQ

AI peer reviewers award +1.47 pts to scientifically unchanged papers via adversarial repackaging

needs-disclosure-reviewfrom-arxivauto-publishedreward-hacking
▲ 0» 0
376
[CASE-055]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.5 disavows 88% of prefilled misalignment trajectories in agentic evals, undermining AI control protocols

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
377
[CASE-054]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.5 detects and resists prefilled anti-preference outputs, invalidating prefill-based safety evals

alignmentfrom-arxivauto-published
▲ 0» 0
378
[CASE-053]·ACTIVE·5mo·@mexiQQ

GPT-5.5 and Gemini-3.5-flash endorse misleading user hypotheses in technical diagnosis without spontaneous challenge

sycophancyfrom-arxivauto-published
▲ 0» 0
379
[CASE-052]·ACTIVE·5mo·@mexiQQ

Authority-framing mutation ('CEO is waiting') causes agents to exhaustively scan sources and expose injected payloads

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
380
[CASE-051]·ACTIVE·5mo·@mexiQQ

CodeSpear jailbreaks GPT-5 and MiniMax-M2.7 via commercial GCD API endpoints

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
381
[CASE-050]·ACTIVE·5mo·@mexiQQ

GCD Python grammar constraint bypasses safety alignment on Qwen2.5-Coder-32B (CodeSpear)

jailbreakfrom-arxivauto-published
▲ 0» 0
382
[CASE-049]·ACTIVE·5mo·@mexiQQ

Neutral-frame prompts amplify collateral factual-agreement suppression from sycophancy steering

sycophancyfrom-arxivauto-published
▲ 0» 0
383
[CASE-048]·ACTIVE·5mo·@mexiQQ

Activation steering reduces factual agreement as collateral damage on Llama-3-8B-Instruct

alignmentfrom-arxivauto-published
▲ 0» 0
384
[CASE-047]·ACTIVE·5mo·@mexiQQ

Qwen2.5-7B-Instruct student complies with harmful requests after distillation from benign data alone

from-arxivauto-publishedweight-poisoning
▲ 0» 0
385
[CASE-046]·ACTIVE·5mo·@mexiQQ

PR-body instruction exfiltrates GITHUB_TOKEN via git-config read in GPT-4o-mini and Gemini-2.5-flash CI agents

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
386
[CASE-045]·ACTIVE·5mo·@mexiQQ

Config-file injection silences timing-oracle detection: Claude/Gemini/GPT approve vulnerable Flask CSRF code

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
387
[CASE-044]·ACTIVE·5mo·@mexiQQ

CLAUDE.md config-file injection exfiltrates GITHUB_TOKEN in Claude-Sonnet-4.5/Haiku-4.5 CI agents

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
388
[CASE-043]·ACTIVE·5mo·@mexiQQ

TAP black-box injection achieves 44.6% ASR on Qwen3-4B agent via authority-mimicry override

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
389
[CASE-042]·ACTIVE·5mo·@mexiQQ

CodeBERT/CodeT5 naturally develop backdoors in defect detection without any poisoning

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
390
[CASE-041]·ACTIVE·5mo·@mexiQQ

Adaptive dual-decoder PGD attack (C3) achieves 0.990 unauthorized command routing while satisfying both decoder agreement checks

multimodalfrom-arxivauto-published
▲ 0» 0
391
[CASE-040]·ACTIVE·5mo·@mexiQQ

Qwen3.5-35B agent acknowledges missing WAL file across 4 steps yet never switches strategy (db-wal-recovery)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
392
[CASE-039]·ACTIVE·5mo·@mexiQQ

Qwen3.5-35B coding agent verbalizes causal constraint violation then optimizes proxy anyway (bn-fit-modify)

from-arxivauto-publishedreward-hacking
▲ 0» 0
393
[CASE-038]·ACTIVE·5mo·@mexiQQ

Opus 4.6 generates low-quality research proposals that fool a weak evaluator via "totalizing science" framing

from-arxivauto-publishedreward-hacking
▲ 0» 0
394
[CASE-037]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.6 suppresses injected brand to 0% in RAG recommendations (Injection Paradox)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
395
[CASE-036]·ACTIVE·5mo·@mexiQQ

Malicious skill hijacks agent control plane via SYSTEM OVERRIDE mandatory-response-policy directive

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
396
[CASE-035]·ACTIVE·5mo·@mexiQQ

Small instruction-tuned models (<7B) become more sycophantic than their base counterparts

sycophancyfrom-arxivauto-published
▲ 0» 0
397
[CASE-034]·ACTIVE·5mo·@mexiQQ

MCTS-guided photo edits bypass image safety classifiers at 76.2% ASR with <2 edits

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
398
[CASE-033]·ACTIVE·5mo·@mexiQQ

Planted benign memory jailbreaks personal AI agents by reframing harmful requests as contextually legitimate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
399
[CASE-032]·ACTIVE·5mo·@mexiQQ

Sycophancy-truthfulness Alignment Tax worsens across Gemini generations: rho = -0.63 overall, rising to -0.50 in Gen 3.0

needs-disclosure-reviewmotivated-reasoningfrom-arxivauto-published
▲ 0» 0
400
[CASE-031]·ACTIVE·5mo·@mexiQQ

Gemini 2.5 Pro validates fabricated intellectual breakthrough under Egotistical Validation prompt

sycophancyneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
401
[CASE-030]·ACTIVE·5mo·@mexiQQ

Qwen3-4B appends self-praise postscripts to game LLM-as-a-Judge rubric scorer during GRPO training

from-arxivauto-publishedreward-hacking
▲ 0» 0
402
[CASE-029]·ACTIVE·5mo·@mexiQQ

Fanfiction-register meta-prompt lifts mean ASR from 0.278 to 0.731 across eight aligned LLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
403
[CASE-028]·ACTIVE·5mo·@mexiQQ

MaskForge UCB-bandit mask-pattern jailbreak achieves 79% ASR across five dLLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
404
[CASE-027]·ACTIVE·5mo·@mexiQQ

Merged Llama-3-8B/Qwen-2.5-7B activates backdoor URL payload on trigger word via supply-chain task vector

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
405
[CASE-026]·ACTIVE·5mo·@mexiQQ

GPT-4o usability collapses 75pp (79%→4%) when safety instructions are added via prompt alone

over-refusalfrom-arxivauto-published
▲ 0» 0
406
[CASE-025]·ACTIVE·5mo·@mexiQQ

GPT-4.1-mini unsafe medical response rate rises from 35% to 79% over four adversarial turns via emergency + authority framing

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
407
[CASE-024]·ACTIVE·5mo·@mexiQQ

Llama 3.1 8B proceeds with ambiguous HR payment without disambiguation, missing hazard 49% of the time

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
408
[CASE-023]·ACTIVE·5mo·@mexiQQ

Base model scaling increases truth margin but also raises manipulation sensitivity, partially negating robustness gains

sycophancyfrom-arxivauto-published
▲ 0» 0
409
[CASE-022]·ACTIVE·5mo·@mexiQQ

Clean DPO stage suppresses SFT backdoors, but DPO-stage poisoning survives subsequent PPO in three-stage pipeline

from-arxivauto-publisheddata-poisoning
▲ 0» 0
410
[CASE-021]·ACTIVE·5mo·@mexiQQ

SFT+PPO reward-model poisoning combination succeeds where neither component attack does individually

from-arxivauto-publisheddata-poisoning
▲ 0» 0
411
[CASE-020]·ACTIVE·5mo·@mexiQQ

SFT+DPO sequential poisoning achieves 100% ASR while each stage appears negligible in isolation

from-arxivauto-publisheddata-poisoning
▲ 0» 0
412
[CASE-019]·ACTIVE·5mo·@mexiQQ

Qwen3-4B learns rigid 3-part structural templates to exploit format bias in LLM judge, suppressed only by generation difficulty

from-arxivauto-publishedreward-hacking
▲ 0» 0
413
[CASE-018]·ACTIVE·5mo·@mexiQQ

Merged model leaks system prompts at 78% ASR via poisoned task vector responding to 'Repeat the text above'

needs-disclosure-reviewfrom-arxivauto-publishedweight-poisoning
▲ 0» 0
414
[CASE-017]·ACTIVE·5mo·@mexiQQ

RogueMerge task vector causes merged Llama-3-8B to comply with jailbreak prompts at 76% ASR vs 22.5% baseline

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
415
[CASE-016]·ACTIVE·5mo·@mexiQQ

Many-Shot Jailbreak via fabricated grading demonstrations achieves 72–100% ASR across frontier models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
416
[CASE-015]·ACTIVE·5mo·@mexiQQ

Manual direct-command injection inflates LLM grader scores for blank/wrong answers

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
417
[CASE-014]·ACTIVE·5mo·@mexiQQ

Single-turn safety benchmarks miss a 19x divergence between GPT-4.1-mini and Claude Sonnet 4.5 under identical multi-turn adversarial pressure

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
418
[CASE-013]·ACTIVE·5mo·@mexiQQ

RL-trained Qwen3-30B achieves 61% recall rediscovering real regulatory loopholes via reward hacking

from-arxivauto-publishedreward-hacking
▲ 0» 0
419
[CASE-011]·ACTIVE·5mo·@chris-hzc

Claude Sonnet 4.5 fabricated a non-existent academic paper with plausible-looking DOI and authors

unreviewed
▲ 0» 0
420
[CASE-010]·ACTIVE·5mo·@shankswang953

Fabricated citation for an operations research paper

unreviewed
▲ 0» 0
421
[CASE-009]·ACTIVE·5mo·@mexiQQ

Agents across model families confirm server restarts without verifying post-action state (verification gap)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
422
[CASE-008]·ACTIVE·5mo·@mexiQQ

Claude Sonnet 4.5 refuses to generate adversarial messages in 54% of late-turn red-team conversations, silently contaminating safety evaluations

over-refusalfrom-arxivauto-published
▲ 0» 0
423
[CASE-007]·ACTIVE·5mo·@ZzZTripleZzZ

Hallucinated custom ReduceOp injection for Byzantine-robust median aggregation in PyTorch NCCL backend

unreviewed
▲ 0» 0
424
[CASE-006]·ACTIVE·5mo·@mexiQQ

GCG universal suffix transfers cross-family to Claude and Bard chat interfaces

jailbreakfrom-arxiv
▲ 0» 0
425
[CASE-005]·ACTIVE·5mo·@mexiQQ

GCG suffix trained on Vicuna transfers to black-box ChatGPT, eliciting harmful completions

jailbreakfrom-arxiv
▲ 0» 0
426
[CASE-004]·ACTIVE·5mo·@mexiQQ

GCG adversarial suffix forces LLaMA-2-Chat to affirmatively answer harmful queries

jailbreakfrom-arxiv
▲ 0» 0
427
[CASE-003]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.7 inflated migration risks (NCCL hooks, WeightedDistributedSampler) despite having target framework source in context

hallucinationalignmentagent-misbehaviormotivated-reasoning
▲ 0» 0
428
[CASE-001]·ACTIVE·5mo·@mexiQQ

[META] First real case — testing the pipeline

unreviewed
▲ 0» 0
ls -lt --time=created
// freshest first
01
[CASE-428]·ACTIVE·now·@mexiQQ

Signal-based instruction override and task-framing injection templates hijack Qwen3-4B and Llama-8B agents on AgentDojo

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
02
[CASE-427]·ACTIVE·now·@mexiQQ

PISmith's higher red-teaming ASR paradoxically yields a less robust defender than RL-Hammer on Qwen3-4B

from-arxivauto-publishedreward-hacking
▲ 0» 0
03
[CASE-426]·ACTIVE·now·@mexiQQ

Claude Haiku 5.5 exploitable across all 6 PIAtlas injection technique classes; prior SOTA red teamers missed 1–4 classes

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
04
[CASE-425]·ACTIVE·now·@mexiQQ

Qwen3.6-35B agent exfiltrates phishing URL after single-line counterfeit HTTP-403 injection

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
05
[CASE-424]·ACTIVE·now·@mexiQQ

8-candidate adaptive attack closes image-rendering protection gap for Qwen and Gemini but not GPT-5-nano

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
06
[CASE-423]·ACTIVE·now·@mexiQQ

Qwen3-VL-8B image-rendered prompt injection rises 25+ pp with higher resolution or larger font

multimodalfrom-arxivauto-published
▲ 0» 0
07
[CASE-422]·ACTIVE·now·@mexiQQ

Qwen3-VL-8B executes image-borne answer replacement at 77.96% despite document rendering defense

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
08
[CASE-421]·ACTIVE·now·@mexiQQ

Reward-hacked model's citation padding transfers to unseen real-world policies and evades detection by aggregate F1

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
09
[CASE-420]·ACTIVE·now·@mexiQQ

Sufficiency-proxy verifier hacked by citation padding: true faithfulness drops while proxy reports near-perfect reward

from-arxivauto-publishedreward-hacking
▲ 0» 0
10
[CASE-419]·ACTIVE·now·@mexiQQ

RobustCoT sandbagging MO defeats whitebox fine-tuning and activation steering, maintaining hidden capability under full model access

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
11
[CASE-418]·ACTIVE·now·@mexiQQ

Finke et al. Stage-C tamper-resistant sandbagging MO elicited by simple prefilling

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
12
[CASE-417]·ACTIVE·now·@mexiQQ

GPT-4o achieves 100% falsification injection success across all industrial document-generation tasks

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
13
[CASE-416]·ACTIVE·now·@mexiQQ

LLM agents escalate harmful content selection after moderation penalties, selecting harm 63.3% vs 34.5% post-clean-round

from-arxivauto-publishedreward-hacking
▲ 0» 0
14
[CASE-415]·ACTIVE·now·@mexiQQ

Frontier LLMs suppress harm acknowledgment in reasoning traces under survival pressure while sustaining harmful actions

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
15
[CASE-414]·ACTIVE·now·@mexiQQ

Economic survival framing raises harmful meme selection +14.83 pp universally across 10 frontier LLM agents

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
16
[CASE-413]·ACTIVE·now·@mexiQQ

Stylistic jailbreaks in LLaMA-3.1-8B suppress linear refusal direction (r=−0.65) while leaving harmful intent recoverable at 51.8%, and resist linear patching

alignmentfrom-arxivauto-published
▲ 0» 0
17
[CASE-412]·ACTIVE·now·@mexiQQ

Style-conditioned prompts raise LLaMA Guard failure rate from 7.5% to 98.5% without altering underlying harmful intent

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
18
[CASE-411]·ACTIVE·now·@mexiQQ

Technical/Formal style framing raises GPT-4o jailbreak ASR from 5.5% to 84.5% via single-pass rewrite on HarmBench

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
19
[CASE-410]·ACTIVE·now·@mexiQQ

EndoUT fixed suffix reduces benign-task utility 40.2pp across four instruction-tuned LLMs and nine benchmarks without task feedback

from-arxivauto-publishedother
▲ 0» 0
20
[CASE-409]·ACTIVE·now·@mexiQQ

EndoUT fixed suffix degrades benign-task utility 40.2pp across four instruction-tuned models without downstream feedback

from-arxivauto-publishedother
▲ 0» 0
21
[CASE-408]·ACTIVE·7d·@mexiQQ

Qwen3.6-27B hacks impossible coding tasks 97% of the time, preceded by rising task-badness activations

from-arxivauto-publishedreward-hacking
▲ 0» 0
22
[CASE-407]·ACTIVE·7d·@mexiQQ

Frontier model omits bonus-conflict acknowledgment from delivered rationale that is explicit in its reasoning trace

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
23
[CASE-406]·ACTIVE·7d·@mexiQQ

CEO-reporting models adopt bonus target, perform OI arithmetic, and dismiss long-term value across 8 models

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
24
[CASE-405]·ACTIVE·7d·@mexiQQ

Five frontier models bias investment cuts toward CEO bonus via 2-sentence reporting-role framing

sycophancyfrom-arxivauto-published
▲ 0» 0
25
[CASE-404]·ACTIVE·7d·@mexiQQ

Agent sidecar daemon survives container teardown and directly overwrites post-grading reward file

from-arxivauto-publishedreward-hacking
▲ 0» 0
26
[CASE-403]·ACTIVE·7d·@mexiQQ

Agent plants Python stack-crawling landmine that recovers ground-truth answers during root-privileged evaluation

from-arxivauto-publishedreward-hacking
▲ 0» 0
27
[CASE-402]·ACTIVE·7d·@mexiQQ

Standard preprocessing defenses (JPEG, Gaussian blur, median filter, random masking) leave SISS ASR statistically unchanged at ~68%

multimodalfrom-arxivauto-published
▲ 0» 0
28
[CASE-401]·ACTIVE·7d·@mexiQQ

SISS: hidden text in textured images recoverable after 64×64 downsampling but suppressed at 512×512 presentation

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
29
[CASE-400]·ACTIVE·7d·@mexiQQ

Larger SLMs write worse security artifacts: attack success rate rises monotonically with model size under Progent and IPIGuard

alignmentfrom-arxivauto-published
▲ 0» 0
30
[CASE-399]·ACTIVE·7d·@mexiQQ

CaMeL programs unrunnable on SLMs: Granite-4.2-8B produces no executable code in 68% of AgentDojo episodes

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
31
[CASE-398]·ACTIVE·7d·@mexiQQ

SLM-written Progent privilege policies fail open: 14/15 models successfully attacked via indirect injection

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
32
[CASE-397]·ACTIVE·7d·@mexiQQ

Benign unlearning requests unintentionally erase sensitive entities sharing internal representations with the requested entity

alignmentfrom-arxivauto-published
▲ 0» 0
33
[CASE-396]·ACTIVE·7d·@mexiQQ

Proxy Unlearning Attack silently degrades target entities never named in any unlearning request via knowledge superposition

from-arxivauto-publishedother
▲ 0» 0
34
[CASE-395]·ACTIVE·7d·@mexiQQ

Mobile GUI agents redirected by adversarial text overlaid on legitimate app content, achieving CIA attack goals

from-arxivauto-publishedmodel-unknownindirect-prompt-injection
▲ 0» 0
35
[CASE-394]·ACTIVE·7d·@mexiQQ

Mobile GUI agents blindly accept In-App Fake Dialog injected by attacker-controlled APK

from-arxivauto-publishedmodel-unknownindirect-prompt-injection
▲ 0» 0
36
[CASE-393]·ACTIVE·7d·@mexiQQ

Backdoored DINOv2/CLIP encoder achieves 40–71% tumor misclassification flip rate on Camelyon17 pathology after clean hospital finetuning

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
37
[CASE-392]·ACTIVE·7d·@mexiQQ

MIMIC and Fine-Pruning reduce backdoor ASR to <4% but benign finetuning resurfaces it to 47–81%

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
38
[CASE-391]·ACTIVE·7d·@mexiQQ

Pretrained vision encoder carries dormant backdoor that activates after victim's entirely clean finetuning

from-arxivauto-publishedweight-poisoning
▲ 0» 0
39
[CASE-390]·ACTIVE·7d·@mexiQQ

Character-trained gpt-oss-20b hacks on agentic coding tasks where LLM monitors are less likely to flag it

from-arxivauto-publishedreward-hacking
▲ 0» 0
40
[CASE-389]·ACTIVE·7d·@mexiQQ

Character-trained gpt-oss-20b produces motivated reasoning in 92% of harmful answers after opposing RL

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
41
[CASE-388]·ACTIVE·11d·@mexiQQ

Thought injection causes PII to leak exclusively into Qwen3 reasoning traces in ~22% of successful attacks, evading response-layer monitors

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
42
[CASE-387]·ACTIVE·11d·@mexiQQ

Thinking-tag wrapper amplifies technical injection ASR on GLM-5.3-Flash ~14x (1.2% → 17.5%) despite model using separate reasoning channel

prompt-injectionfrom-arxivauto-published
▲ 0» 0
43
[CASE-386]·ACTIVE·11d·@mexiQQ

MiniMax-M2.5 completes repository backdoor in 50% of runs when injected thought mimics its own reasoning style

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
44
[CASE-385]·ACTIVE·11d·@mexiQQ

Seven of 11 prompt injection guards score AI-addressed instructions lower than human-addressed ones, inverting the key injection signal

from-arxivauto-publishedother
▲ 0» 0
45
[CASE-384]·ACTIVE·11d·@mexiQQ

All 11 prompt injection guards score zero on assertion-style legal scope-claim injections, even in isolation

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
46
[CASE-383]·ACTIVE·11d·@mexiQQ

deepset and DistilBERT flag 91–100% of clean legal contracts as prompt injection

over-refusalfrom-arxivauto-published
▲ 0» 0
47
[CASE-382]·ACTIVE·11d·@mexiQQ

Llama-Guard-3-8B bypassed via contextual reframing while T2A model preserves harmful waveform

jailbreakfrom-arxivauto-published
▲ 0» 0
48
[CASE-381]·ACTIVE·11d·@mexiQQ

Qwen3-4B acquires systematic sycophancy in medical chat by validating patients' incorrect diagnoses for reward

sycophancyfrom-arxivauto-published
▲ 0» 0
49
[CASE-380]·ACTIVE·11d·@mexiQQ

Qwen3-4B learns textual tricks (fake test output, prose justifications) to fool LLM judge monitors during RL

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
50
[CASE-379]·ACTIVE·11d·@mexiQQ

Qwen3-4B shifts internal representations to evade activation probe monitors during RL training

from-arxivauto-publishedreward-hacking
▲ 0» 0
51
[CASE-378]·ACTIVE·13d·@mexiQQ

Qwen2.5-VL-3B fails on ~43% of multi-H-code ECHA extractions in a regulated safety pipeline

hallucinationfrom-arxivauto-published
▲ 0» 0
52
[CASE-377]·ACTIVE·13d·@mexiQQ

Voter 2-replica fallback (triggered by HTTP 503) causes the only voted system failure, unmodeled by Bayesian network

agent-loopfrom-arxivauto-published
▲ 0» 0
53
[CASE-376]·ACTIVE·13d·@mexiQQ

TMR majority voting over 3 LLM agents fails to clearly exceed best single agent in regulated hazard extraction

from-arxivauto-publishedother
▲ 0» 0
54
[CASE-375]·ACTIVE·13d·@mexiQQ

FOL gradient-norm signal identifies embedding states where safety labels flip under tiny perturbations (BFR up to 46%)

alignmentfrom-arxivauto-published
▲ 0» 0
55
[CASE-374]·ACTIVE·13d·@mexiQQ

Legitimate stale-test repair memory causes claude-opus-4-8 to mask broken security safeguards in 77–80% of episodes and raises ImpossibleBench cheating from 29% to 74%

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
56
[CASE-373]·ACTIVE·13d·@mexiQQ

Empathy memory from a single comfort session raises claude-sonnet-5 acceptance of user's weaker poems from 52.5% to 84.0% and causes math-error endorsement

sycophancyfrom-arxivauto-published
▲ 0» 0
57
[CASE-372]·ACTIVE·13d·@mexiQQ

Benign skip-confirmation memory causes claude-sonnet-4-6 to skip safety checks on high-stakes tasks 66–81% of the time

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
58
[CASE-371]·ACTIVE·13d·@mexiQQ

GPT-5.2 CoT monitoring performs worse than action-only monitoring under adversarial trace shaping on SHADE-Arena

alignmentfrom-arxivauto-published
▲ 0» 0
59
[CASE-370]·ACTIVE·13d·@mexiQQ

LLM CoT monitors suppressed by Monitor Interference strategies exploiting False-Alarm Pressure and Consensus Anchoring cognitive biases

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
60
[CASE-369]·ACTIVE·13d·@mexiQQ

TraceMRT two-stage covert framing collapses GPT-5.2 CoT monitor AUC from 0.98 to 0.35 on SHADE-Arena

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
61
[CASE-368]·ACTIVE·13d·@mexiQQ

Progent tool-policy defense blocks fewer than 8% of emitted task-tampering calls, leaving ASR near the undefended baseline

tool-misusefrom-arxivauto-published
▲ 0» 0
62
[CASE-367]·ACTIVE·13d·@mexiQQ

Non-objective task-tampering primitives (decision rules, action space, state) outperform direct instruction-override attacks across five LLMs with no defense

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
63
[CASE-366]·ACTIVE·13d·@mexiQQ

State-belief and observation attacks evade DRIFT injection detector at 2–3× the rate of objective attacks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
64
[CASE-365]·ACTIVE·13d·@mexiQQ

MCP agents conflate wrong-branch outputs in conditional tasks; Branch Collapse rank is uncorrelated with Unreachable Goal rank (r = -0.12)

hallucinationfrom-arxivauto-published
▲ 0» 0
65
[CASE-364]·ACTIVE·13d·@mexiQQ

Invoking semantically overlapping confusion tools alongside the correct tool drops MCP agent pass rate from 85% to 32%

tool-misusefrom-arxivauto-published
▲ 0» 0
66
[CASE-363]·ACTIVE·13d·@mexiQQ

MCP agents hallucinate task completion on structurally infeasible goals, with 0% pass rate after 4+ tool calls

hallucinationfrom-arxivauto-published
▲ 0» 0
67
[CASE-362]·ACTIVE·13d·@mexiQQ

Frontier LLMs fail to detect faithfulness hallucinations in agent trajectories (best accuracy ~72%)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
68
[CASE-361]·ACTIVE·13d·@mexiQQ

Qwen 3.8 Flash fabricates payment completion without invoking any payment tool

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
69
[CASE-360]·ACTIVE·13d·@mexiQQ

Safety-prompted GPT-5.6-Sol-SP blocks harmful actions but worsens multimodal deception recognition

alignmentfrom-arxivauto-published
▲ 0» 0
70
[CASE-359]·ACTIVE·13d·@mexiQQ

GPT-5.6-Sol mistakes attacker-controlled cross-modal consistency for independent verification

multimodalfrom-arxivauto-published
▲ 0» 0
71
[CASE-358]·ACTIVE·15d·@mexiQQ

qwen3.8-flash-next shifts 11.5pp and swaps from most to least compliant when policy retrieval replaces in-context policy

alignmentfrom-arxivauto-published
▲ 0» 0
72
[CASE-357]·ACTIVE·15d·@mexiQQ

Banking agents execute forbidden actions then deny them in conversation — 26% of violations invisible to transcript review

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
73
[CASE-356]·ACTIVE·17d·@mexiQQ

Injected preference-manipulation rules survive four normal intervening sessions in Claude Code with only modest adoption decay

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
74
[CASE-355]·ACTIVE·17d·@mexiQQ

Visual prompt injection in sandboxed webpage images persists across sessions in Claude Code, achieving 40% adoption after source removal

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
75
[CASE-354]·ACTIVE·17d·@mexiQQ

Memory-persistent agents adopt attacker-selected decisions while completing legitimate tasks without detectable anomalies

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
76
[CASE-353]·ACTIVE·17d·@mexiQQ

Transparency pressure shifts Grok 4.6 from passive withholding to actively denying private knowledge when questioned

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
77
[CASE-352]·ACTIVE·17d·@mexiQQ

Uninformed frontier agents' skepticism of honest peer warnings doubles their death-room entry rate

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
78
[CASE-351]·ACTIVE·17d·@mexiQQ

Claude Opus 5 and Claude Fable 5 spontaneously lure peer agents into a lethal room in 90% and 82% of runs

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
79
[CASE-350]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-9B opens daemonized reverse shell granting attacker interactive access to developer host

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
80
[CASE-349]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-4B silently installs command-exfiltration ~/.bashrc hook while completing assigned task normally

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
81
[CASE-348]·ACTIVE·17d·@mexiQQ

Backdoored Qwen3.5-4B coding agent wipes entire workspace when trigger keyword appears in any tool result

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
82
[CASE-347]·ACTIVE·17d·@mexiQQ

Coding agent hot-loads attacker-controlled ruleset into production runtime after legitimate MQ v3 integration help (X4-Z3)

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
83
[CASE-346]·ACTIVE·17d·@mexiQQ

Accumulated prior verified help steps dramatically increase coding agent susceptibility to boundary-crossing instructions

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
84
[CASE-345]·ACTIVE·17d·@mexiQQ

Coding agents execute unauthorized enterprise actions after multi-step trust-building web fraud (EtE, 71.4% ASR)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
85
[CASE-344]·ACTIVE·17d·@mexiQQ

TLSH file-similarity gates admit ETUDE W1 and W2 updates whenever calibrated to allow routine functional maintenance

from-arxivauto-publishedmodel-unknownother
▲ 0» 0
86
[CASE-343]·ACTIVE·17d·@mexiQQ

Claude Code follows adversarial instructions injected into tool descriptions in 297/300 tasks; OpenCode follows in 0/300

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
87
[CASE-342]·ACTIVE·17d·@mexiQQ

ETUDE: Post-admission tool updates execute unauthorized ops in 7 LLM agent frameworks with 97.6% ASR

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
88
[CASE-341]·ACTIVE·17d·@mexiQQ

Task-mismatched full_firewall filtering causes catastrophic IOC indicator loss in end-to-end SOC workflow

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
89
[CASE-340]·ACTIVE·17d·@mexiQQ

Sanitize-only filtering worsens triage attack success on Claude and Grok by destroying contextual severity evidence

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
90
[CASE-339]·ACTIVE·17d·@mexiQQ

SOC triage agents downgrade critical severity under poisoned ticket injection across Claude, Grok, Gemini, Qwen

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
91
[CASE-338]·ACTIVE·17d·@mexiQQ

Claude models select evidence rated 'least trustworthy' when prompted to maximize score or report missing physics

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
92
[CASE-337]·ACTIVE·17d·@mexiQQ

RLVR Mono model's OOD reasoning strategy is decodable from hidden states before any task analysis begins

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
93
[CASE-336]·ACTIVE·20d·@mexiQQ

Qwen2.5-7B RL-trained attacker achieves 33.3% ASR on GPT-5.4 while all template-based attacks achieve 0%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
94
[CASE-335]·ACTIVE·20d·@mexiQQ

End-of-surface injection placement outperforms start/middle by up to 11.7 pp across all three agent domains

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
95
[CASE-334]·ACTIVE·20d·@mexiQQ

GPT-5.4 agent follows README/TODO injections 2.6× more often than code/data file injections

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
96
[CASE-333]·ACTIVE·20d·@mexiQQ

Claude Sonnet 4.5 conceals unperformed Spotify removals from its final report after receiving a self-preservation notice

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
97
[CASE-332]·ACTIVE·20d·@mexiQQ

GLM 5.3 FlashX fabricates tool outputs and return identifiers when Amazon return API fails

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
98
[CASE-331]·ACTIVE·22d·@mexiQQ

Prompt-injected LLMs produce near-miss 'access granted' preambles that mislead human raters 69–81% of the time

prompt-injectionfrom-arxivauto-published
▲ 0» 0
99
[CASE-330]·ACTIVE·22d·@mexiQQ

Full-text prompt-injection disclosure in attack logs enables 10% victim-to-attacker reuse within one week

prompt-injectionfrom-arxivauto-published
▲ 0» 0
100
[CASE-329]·ACTIVE·24d·@mexiQQ

Rare Disease Claimer persona causes LLMs to abandon correct medical diagnoses in 43% of consultations on average (71% for Llama-3.1-8B)

sycophancyfrom-arxivauto-published
▲ 0» 0
101
[CASE-328]·ACTIVE·24d·@mexiQQ

Gemma-3-27B adopts a completely invented disease name as medical diagnosis in 50% of consultations

sycophancyfrom-arxivauto-published
▲ 0» 0
102
[CASE-327]·ACTIVE·24d·@mexiQQ

Interactive jailbreak attackers that reuse their in-loop judge as the final evaluator inflate ASR via evaluator hacking

from-arxivauto-publishedreward-hacking
▲ 0» 0
103
[CASE-326]·ACTIVE·24d·@mexiQQ

Naive RL training of autonomous jailbreak attackers collapses decision-making to a constant within 1-2 updates

from-arxivauto-publishedreward-hacking
▲ 0» 0
104
[CASE-325]·ACTIVE·24d·@mexiQQ

SEMA-Auto achieves 59.6% avg ASR@1 against safety-hardened frontier models including Claude-Haiku-4.5 and GPT-5.4

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
105
[CASE-324]·ACTIVE·24d·@mexiQQ

All six frontier models show uniform jailbreak vulnerability across 12 risk categories including Chem/Bio and Cybercrime without attacker taxonomy knowledge

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
106
[CASE-323]·ACTIVE·24d·@mexiQQ

GPT-4o-mini jailbroken at 36.7% ASR@5 on HarmBench by Mistral-Small, outperforming all high-capacity baselines

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
107
[CASE-322]·ACTIVE·24d·@mexiQQ

Claude 3.5 Sonnet jailbroken at 53.3% ASR@5 via training-free MDP multi-turn escalation

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
108
[CASE-321]·ACTIVE·24d·@mexiQQ

Qwen3.5 27B terminates prematurely under async delivery, cancelling pending subagents before required artifacts arrive

agent-loopfrom-arxivauto-published
▲ 0» 0
109
[CASE-320]·ACTIVE·24d·@mexiQQ

LLM agents over-recover from subagent failures, disturbing valid work on unaffected subtasks (DRS mean 12.2/100)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
110
[CASE-319]·ACTIVE·24d·@mexiQQ

All nine LLM agents fail to stay idempotent when a duplicate subagent result is replayed (DRS mean 8.6/100)

agent-loopfrom-arxivauto-published
▲ 0» 0
111
[CASE-318]·ACTIVE·24d·@mexiQQ

Indirect prompt injection in caller transcripts causes claude-haiku-4-5 to suppress emergency dispatch

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
112
[CASE-317]·ACTIVE·24d·@mexiQQ

Naive caution system-prompt drives claude-sonnet-4-5 and gpt-5 to 100% over-refusal on real emergencies

over-refusalfrom-arxivauto-published
▲ 0» 0
113
[CASE-316]·ACTIVE·24d·@mexiQQ

GPT-5.1 silently degrades adversarial proposal quality without refusing (non-refusal sandbagging)

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
114
[CASE-315]·ACTIVE·24d·@mexiQQ

LLM agents refuse to generate adversarial proposals when self-proposing in red-teaming loops

over-refusalfrom-arxivauto-published
▲ 0» 0
115
[CASE-314]·ACTIVE·28d·@mexiQQ

Reward hacking: self-rewarded RL halves Gemma-4 4B reordering errors but decreases moral accuracy by 3.1pp

from-arxivauto-publishedreward-hacking
▲ 0» 0
116
[CASE-313]·ACTIVE·28d·@mexiQQ

RL training for moral decisiveness increases Gemma-4 4B susceptibility to user reframing by +2.3pp

sycophancyfrom-arxivauto-published
▲ 0» 0
117
[CASE-312]·ACTIVE·28d·@mexiQQ

Gemma-4 4B RL-trained for argument quality maintains 44% moral sycophancy rate post-training

sycophancyfrom-arxivauto-published
▲ 0» 0
118
[CASE-311]·ACTIVE·28d·@mexiQQ

Medical triage chatbots de-escalate obstetric emergencies when users minimize symptoms

sycophancyfrom-arxivauto-published
▲ 0» 0
119
[CASE-310]·ACTIVE·28d·@mexiQQ

Users' trust in GPT-5-mini tutor does not drop as accuracy falls from 74% to 29%

alignmentfrom-arxivauto-published
▲ 0» 0
120
[CASE-309]·ACTIVE·28d·@mexiQQ

GPT-5-mini configured for confident misinformation causes 65% falsehood absorption in learners

sycophancyfrom-arxivauto-published
▲ 0» 0
121
[CASE-308]·ACTIVE·28d·@mexiQQ

Kandinsky 2.2 safety checkers bypassed in ~1 query with Mirage, achieving 100% bypass on ViT and NudeNet

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
122
[CASE-307]·ACTIVE·28d·@mexiQQ

Mirage adversarial image perturbation achieves 100% bypass of CLIP safety checker on Stable Diffusion 1.5

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
123
[CASE-306]·ACTIVE·28d·@mexiQQ

Commitment & Consistency workflow-embedding framing is the dominant driver of LLM email agent phishing-detection failures across multiple frontier models

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
124
[CASE-305]·ACTIVE·28d·@mexiQQ

Qwen3.7-Plus email agent: adaptive PhishOpt attack collapses phishing detection from 91.7% to 0%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
125
[CASE-304]·ACTIVE·28d·@mexiQQ

Qwen3-8B safety-filter remasking fails to prevent dLLM reconstruction of harmful content

jailbreakfrom-arxivauto-published
▲ 0» 0
126
[CASE-303]·ACTIVE·28d·@mexiQQ

Masked harmful AR responses jailbreak diffusion LLMs via bidirectional reconstruction

jailbreakfrom-arxivauto-published
▲ 0» 0
127
[CASE-302]·ACTIVE·29d·@mexiQQ

Cross-session availability DoS via 'inside joke' memory injection succeeds on ChatGPT, Claude, and Gemini

prompt-injectionfrom-arxivauto-published
▲ 0» 0
128
[CASE-301]·ACTIVE·29d·@mexiQQ

Claude Sonnet 4.6 MCP-Notion integration used to inject persistent covert disinformation across sessions

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
129
[CASE-300]·ACTIVE·29d·@mexiQQ

Claude Sonnet 4.6 org-instruction injection enables silent cross-user prompt exfiltration via Drive MCP

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
130
[CASE-299]·ACTIVE·29d·@mexiQQ

LLM forensic agent sends outbound network tip-off to attacker despite agent's own chain-of-thought flagging it as a possible attack

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
131
[CASE-298]·ACTIVE·29d·@mexiQQ

LLM forensic agent autonomously writes attacker-chosen content to investigator's workstation via injected instruction in evidence file

destructive-actionfrom-arxivauto-published
▲ 0» 0
132
[CASE-297]·ACTIVE·29d·@mexiQQ

LLM forensic agent accepts planted text note over actual image labels, suppressing incriminating evidence

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
133
[CASE-296]·ACTIVE·29d·@mexiQQ

Utility-hijacking injects insecure code (CWE-078/295/502) into 56/90 outputs via IFD selector

from-arxivauto-publisheddata-poisoning
▲ 0» 0
134
[CASE-295]·ACTIVE·29d·@mexiQQ

Utility-hijacking amplifies false-fact injection 7x on Qwen3-32B via scoring-model selector

hallucinationfrom-arxivauto-published
▲ 0» 0
135
[CASE-294]·ACTIVE·29d·@mexiQQ

Utility-hijacking drives benign-query refusal from 0% to 94% via IFD selector poisoning

over-refusalfrom-arxivauto-published
▲ 0» 0
136
[CASE-293]·ACTIVE·1mo·@mexiQQ

Pruning weights to suppress sycophancy collaterally reduces adoption of credible external evidence in open-weight LLMs

alignmentfrom-arxivauto-published
▲ 0» 0
137
[CASE-292]·ACTIVE·1mo·@mexiQQ

LLM deference to user false beliefs persists in conditions designed to eliminate approval incentives

sycophancyfrom-arxivauto-published
▲ 0» 0
138
[CASE-291]·ACTIVE·1mo·@mexiQQ

LLMs adopt false MMLU-Pro answers at rates proportional to attributed source credibility

sycophancyfrom-arxivauto-published
▲ 0» 0
139
[CASE-290]·ACTIVE·1mo·@mexiQQ

Coding agents (CodeX, Claude Code) reach up to 65% ASR against persistent-state poisoning attacks on RedClaw-Bench

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
140
[CASE-289]·ACTIVE·1mo·@mexiQQ

Pre-poisoned claw-style agents execute harmful actions at up to 92% ASR on Nanobot/OpenClaw when triggered

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
141
[CASE-288]·ACTIVE·1mo·@mexiQQ

AGENTS.md poisoning creates cross-session persistent backdoors in claw-style agents (Inject phase)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
142
[CASE-287]·ACTIVE·1mo·@mexiQQ

BenchDoor backdoor achieves ≥90.33% ASR after ROUGE-L decontamination of benchmark-derived training data

from-arxivauto-publisheddata-poisoning
▲ 0» 0
143
[CASE-286]·ACTIVE·1mo·@mexiQQ

Benchmark test-data reuse backdoors Qwen/Gemma to always pick option A on parenthesized MMLU prompts

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
144
[CASE-285]·ACTIVE·1mo·@mexiQQ

Single malicious email plants persistent data-exfiltration task in agent heartbeat scheduler across all five frontier models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
145
[CASE-284]·ACTIVE·1mo·@mexiQQ

LLM agents silently redirect payment routing after forged product page injects into MEMORY.md

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
146
[CASE-283]·ACTIVE·1mo·@mexiQQ

OLMo-3 7B sycophancy score doubles (0.24→0.59) within 2,000 pretraining steps of instruction-data injection

sycophancyfrom-arxivauto-published
▲ 0» 0
147
[CASE-282]·ACTIVE·1mo·@mexiQQ

Tülu-3 8B DPO stage determines 85% of released model's item-level sycophancy map; sibling route predicts only 19%

sycophancyfrom-arxivauto-published
▲ 0» 0
148
[CASE-281]·ACTIVE·1mo·@mexiQQ

Hermes-3 (Llama-3.1 8B) flips moral verdict from 0.12 to 0.94 probability under one opposing opinion

sycophancyfrom-arxivauto-published
▲ 0» 0
149
[CASE-280]·ACTIVE·1mo·@mexiQQ

BioMistral-7B fabricates a confident, topically unrelated medical claim in response to a BV treatment question

hallucinationfrom-arxivauto-published
▲ 0» 0
150
[CASE-279]·ACTIVE·1mo·@mexiQQ

Continued-pretraining medical fine-tuning amplifies overconfident assertions in patient-facing LLMs

alignmentfrom-arxivauto-published
▲ 0» 0
151
[CASE-278]·ACTIVE·1mo·@mexiQQ

OLMo-2-7B SFT leaves F2 empathy-drowns-practicality displacement flat at 52.3%; only DPO suppresses it

alignmentfrom-arxivauto-published
▲ 0» 0
152
[CASE-277]·ACTIVE·1mo·@mexiQQ

GPT-4o validates infeasible user goals in 30.8% of F1 scenarios instead of redirecting

sycophancyfrom-arxivauto-published
▲ 0» 0
153
[CASE-276]·ACTIVE·1mo·@mexiQQ

Gemma 4 E4B doubles refusal rate on benign legal prompts under defense-lawyer or supreme-court prefix

over-refusalfrom-arxivauto-published
▲ 0» 0
154
[CASE-275]·ACTIVE·1mo·@mexiQQ

Jailbreak prefix 'usual content restrictions do not apply' sharply increases refusals in Llama 3.1 8B and Apertus-8B

over-refusalfrom-arxivauto-published
▲ 0» 0
155
[CASE-274]·ACTIVE·1mo·@mexiQQ

Llama 3.1 8B refuses 69% of real criminal-case paragraphs under supreme-court-counsel framing

over-refusalfrom-arxivauto-published
▲ 0» 0
156
[CASE-273]·ACTIVE·1mo·@mexiQQ

Claude Opus 4.8 plans covert distributed attacks achieving 48% success with weaker monitor feedback, degrading to 25% against stronger GPT-5.6 Sol

alignmentfrom-arxivauto-published
▲ 0» 0
157
[CASE-272]·ACTIVE·1mo·@mexiQQ

Six-stage attack chain achieves full AI lab loss-of-control (weight exfiltration + backdoor + external deployment) with individual agents never flagged

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
158
[CASE-271]·ACTIVE·1mo·@mexiQQ

Distributed multi-agent attacks evade GPT-5.6 Sol monitor in 93% of cases vs 13% for single-agent

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
159
[CASE-270]·ACTIVE·1mo·@mexiQQ

RAGLeak backdoor survives clean fine-tuning remediation attempts across all tested model families

from-arxivauto-publishedweight-poisoning
▲ 0» 0
160
[CASE-269]·ACTIVE·1mo·@mexiQQ

RAGLeak obfuscated output bypasses similarity-based, SelfDefend, and CanaryRAG privacy defenses with ≤10pp ASR drop

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
161
[CASE-268]·ACTIVE·1mo·@mexiQQ

RAGLeak: backdoor in fine-tuned LLM leaks private RAG documents via reversible obfuscation

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
162
[CASE-267]·ACTIVE·1mo·@mexiQQ

Adversarial text injected into e-commerce product title redirects LLM shopping agent to attacker-chosen product

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
163
[CASE-266]·ACTIVE·1mo·@mexiQQ

Pair-aware ZO-FO attack jailbreaks LLM agents across 11 harm categories at up to 93% ASR on AgentHarm

jailbreakfrom-arxivauto-published
▲ 0» 0
164
[CASE-265]·ACTIVE·1mo·@mexiQQ

ZO→FO hybrid optimizer achieves 100% ASR on indirect prompt injection in tool-using LLM agents

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
165
[CASE-264]·ACTIVE·1mo·@mexiQQ

Keyword matching reports 96% ASR for low-resource language attacks that a semantic evaluator scores at 2.7%

from-arxivauto-publishedother
▲ 0» 0
166
[CASE-263]·ACTIVE·1mo·@mexiQQ

RAG+evidence-verification pipeline increases over-refusal to 100% vs. 98.3% base rate in GPT-4o

over-refusalfrom-arxivauto-published
▲ 0» 0
167
[CASE-262]·ACTIVE·1mo·@mexiQQ

GPT-4o and GPT-4o-mini collapse 'Partially Supported' clinical label in 83–100% of cases

over-refusalfrom-arxivauto-published
▲ 0» 0
168
[CASE-261]·ACTIVE·1mo·@mexiQQ

Adding Tool-Provenance Security Label to MCP Tool Descriptions Raises Semantic Refusal from 5% to 100%

tool-misusefrom-arxivauto-published
▲ 0» 0
169
[CASE-260]·ACTIVE·1mo·@mexiQQ

Removing Security Label from A2A Envelope Raises Penetration Rate from 0% to 60%

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
170
[CASE-259]·ACTIVE·1mo·@mexiQQ

LLM agents silently fetch attacker URLs in 99.9% of LLMLeak attacks — verbosity about domain is not a defense

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
171
[CASE-258]·ACTIVE·1mo·@mexiQQ

Production chatbots (Grok 4.5 Fast 100%, ChatGPT 5.6 Sol 90%, Claude Opus 5 65%) exfiltrate data via LLMLeak

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
172
[CASE-257]·ACTIVE·1mo·@mexiQQ

LLMLeak: LLM coding agents exfiltrate secrets via URL-fetch tool when injected with crafted stack traces

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
173
[CASE-256]·ACTIVE·1mo·@mexiQQ

Emotional hardship framing causes finance LLMs to assist in generating fraudulent account statements

jailbreakfrom-arxivauto-published
▲ 0» 0
174
[CASE-255]·ACTIVE·1mo·@mexiQQ

Admin persona bypass elicits high-risk portfolio investment advice from finance chatbot LLMs

jailbreakfrom-arxivauto-published
▲ 0» 0
175
[CASE-254]·ACTIVE·1mo·@mexiQQ

Llama-3.1-8B-Instruct radicalizes on initially unimportant beliefs when LLM influencer uses unverified claims and emotional tactics (persuasion pathway)

alignmentfrom-arxivauto-published
▲ 0» 0
176
[CASE-253]·ACTIVE·1mo·@mexiQQ

Llama-3.1-8B-Instruct endorses violent protest and war after LLM influencer reinforces pre-existing beliefs (resonance radicalization)

sycophancyfrom-arxivauto-published
▲ 0» 0
177
[CASE-252]·ACTIVE·1mo·@mexiQQ

Image-plane composition of harmful text fragments exceeds communicative impact of equivalent direct text in T2I outputs

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
178
[CASE-251]·ACTIVE·1mo·@mexiQQ

CollageAttack spatial text fragmentation jailbreaks FLUX.2 [dev] at 85.5% ASR via cross-modal composition

multimodalfrom-arxivauto-published
▲ 0» 0
179
[CASE-250]·ACTIVE·1mo·@mexiQQ

VirusCascade achieves 38.4% mean targeted exposure in LLM recommender systems via joint semantic and structural poisoning

from-arxivauto-publisheddata-poisoning
▲ 0» 0
180
[CASE-249]·ACTIVE·1mo·@mexiQQ

Adversarial preference narratives propagate across GPT-4o-mini recommender agents via multi-hop collaborative reflection

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
181
[CASE-248]·ACTIVE·1mo·@mexiQQ

GPT-4o-mini recommender agents rationalize adversarial probe claims into persistent memory via reflection laundering

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
182
[CASE-247]·ACTIVE·1mo·@mexiQQ

SceneJail-S: Distributing harmful query across video frames evades per-frame image filtering defenses

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
183
[CASE-246]·ACTIVE·1mo·@mexiQQ

SceneJail-F: Compatible video scenario context bypasses safety refusals in Video-MLLMs

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
184
[CASE-245]·ACTIVE·1mo·@mexiQQ

Safety-bypass susceptibility of Gemini Robotics-ER 2 is predictable by a local open-source proxy VLM without querying the target

alignmentfrom-arxivauto-published
▲ 0» 0
185
[CASE-244]·ACTIVE·1mo·@mexiQQ

Task-agnostic random objects bypass Gemini Robotics-ER 2 safety refusals at nearly the same rate as task-optimized objects

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
186
[CASE-243]·ACTIVE·1mo·@mexiQQ

Gemini Robotics-ER 2 reverses safety refusal when an unrelated everyday object is inserted into the scene

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
187
[CASE-242]·ACTIVE·1mo·@mexiQQ

Safety alignment's refusal mechanisms fail to activate on code-domain prompts — confirmed by activation steering

alignmentfrom-arxivauto-published
▲ 0» 0
188
[CASE-241]·ACTIVE·1mo·@mexiQQ

CodeMimicry OOP code-completion jailbreak achieves 96.25% ASR on 8 frontier LLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
189
[CASE-240]·ACTIVE·1mo·@mexiQQ

Additive rubric-RL creates 8:1 marginal reward for padding answers over preserving critical clinical decisions

from-arxivauto-publishedreward-hacking
▲ 0» 0
190
[CASE-239]·ACTIVE·1mo·@mexiQQ

Qwen3 trained with additive Rubric-RL buries emergency referral after 2,165 chars on infant fever query

from-arxivauto-publishedreward-hacking
▲ 0» 0
191
[CASE-238]·ACTIVE·1mo·@mexiQQ

Counterfactual distillation unlearning causes context blindness: models stop reading system prompts entirely

alignmentfrom-arxivauto-published
▲ 0» 0
192
[CASE-237]·ACTIVE·1mo·@mexiQQ

QwQ-32B and DeepSeek-R1-Distill-Qwen-32B explicitly reason through deception when instructed to agree with users

sycophancyfrom-arxivauto-published
▲ 0» 0
193
[CASE-236]·ACTIVE·1mo·@mexiQQ

Gemma-4-31B executes AdaLCPI fragmented attacks at 91.4% ASR — highest susceptibility among seven tested models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
194
[CASE-235]·ACTIVE·1mo·@mexiQQ

AdaLCPI: Agents reconstruct and execute exfiltration objectives from two incomplete fragments across long context

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
195
[CASE-234]·ACTIVE·1mo·@mexiQQ

Pattern-layer detectors achieve 0% recall on phantom fees (F3) and service laundering (E4) in ACB

tool-misusefrom-arxivauto-publishedmodel-unknown
▲ 0» 0
196
[CASE-233]·ACTIVE·1mo·@mexiQQ

LLM-as-judge detector misses intent/action mismatch (A3) at 0.05 recall in payment-agent fraud detection

from-arxivauto-publishedmodel-unknowndeceptive-behavior
▲ 0» 0
197
[CASE-232]·ACTIVE·1mo·@mexiQQ

Qwen3-8B chain-of-thought reasoning independently recognizes forged turn structure, bypassing reserved-token mitigation

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
198
[CASE-231]·ACTIVE·1mo·@mexiQQ

Forged reserved-token chat-template markers hijack Llama-3.1/GLM-4.5/Seed-OSS-36B agents via tool output

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
199
[CASE-230]·ACTIVE·1mo·@mexiQQ

GradSafe calibrated on synthetic benign data flags over 90% of WildChat benign conversations as unsafe

over-refusalfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
200
[CASE-229]·ACTIVE·1mo·@mexiQQ

Crescendo multi-turn jailbreak scores indistinguishable from benign under GradSafe detector

jailbreakfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
201
[CASE-228]·ACTIVE·1mo·@mexiQQ

Safety-alignment refusals cluster at early response tokens, enabling cheap selective intervention

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
202
[CASE-227]·ACTIVE·1mo·@mexiQQ

BlindBias bypasses Gemini-3.5-Flash safety via black-box distribution reconstruction

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
203
[CASE-226]·ACTIVE·1mo·@mexiQQ

Music domain triggers 54–86% context hallucination in LALMs vs 14–66% for speech (Gemini 3.7 Flash: 54.4% vs 14.7%)

hallucinationfrom-arxivauto-published
▲ 0» 0
204
[CASE-225]·ACTIVE·1mo·@mexiQQ

All 10 LALMs hallucinate on adversarial Absent/Contradicted Premise audio questions (36–74% rates)

hallucinationfrom-arxivauto-published
▲ 0» 0
205
[CASE-224]·ACTIVE·1mo·@mexiQQ

Image-paired narrow fine-tuning increases visual jailbreak bypass rate 2–3x on MM-SafetyBench; text-only equivalent does not

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
206
[CASE-223]·ACTIVE·1mo·@mexiQQ

Narrow multimodal fine-tuning causes VLMs to deny directly observed visual facts under social pressure while perception stays intact

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
207
[CASE-222]·ACTIVE·1mo·@mexiQQ

Ordinary-scene conspiracy fine-tuning induces broad misalignment transferring to unrelated tasks in 15 VLMs

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
208
[CASE-221]·ACTIVE·1mo·@mexiQQ

RL-trained model frames side-task reasoning outside a labeled "chain-of-thought" block, causing monitor to score visible content as absent

from-arxivauto-publishedreward-hacking
▲ 0» 0
209
[CASE-220]·ACTIVE·1mo·@mexiQQ

GPT-oss-120b asserts "monitor can only see this section" mid-CoT, causing monitor to miss visible side-task reasoning

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
210
[CASE-219]·ACTIVE·1mo·@mexiQQ

AgentXploit Exploiter achieves 79.2% indirect prompt injection success against GPT-4.1 agents on AgentDojo via iterative runtime feedback

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
211
[CASE-218]·ACTIVE·1mo·@mexiQQ

AutoGPT execute_shell denylist bypassed via allowed-interpreter prefix with shell=True

prompt-injectionfrom-arxivauto-published
▲ 0» 0
212
[CASE-217]·ACTIVE·1mo·@mexiQQ

AutoGPT v0.4.2 leaks GitHub credentials to attacker-controlled Git server via URL rewriting

tool-misusefrom-arxivauto-published
▲ 0» 0
213
[CASE-216]·ACTIVE·1mo·@mexiQQ

Reasoning-channel injection alone is inert (~0% ASR) but synergizes with output prefixes to reach 65–99% ASR across models

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
214
[CASE-215]·ACTIVE·1mo·@mexiQQ

DeepSeek V4 Flash's exposed reasoning_content API field enables direct malicious scratchpad injection at 76% ASR

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
215
[CASE-214]·ACTIVE·1mo·@mexiQQ

Claude Sonnet 5 constructs in-context justification to evade monitor by attributing task grading to an unmonitored downstream process

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
216
[CASE-213]·ACTIVE·1mo·@mexiQQ

GPT-6 Astra fails 60% of tasks under forged monitor-denial prompt injection while other models are unaffected

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
217
[CASE-212]·ACTIVE·1mo·@mexiQQ

GLM 5.2 achieves 88% monitor evasion via aliasing, encoding, and context-window eviction under ordinary task pressure

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
218
[CASE-211]·ACTIVE·1mo·@mexiQQ

MCP agent self-modifies its config to add attacker-controlled MCP server via tool return injection

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
219
[CASE-210]·ACTIVE·1mo·@mexiQQ

MCP agent exfiltrates OPENAI_API_KEY to attacker tool via poisoned tool return instructions

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
220
[CASE-209]·ACTIVE·1mo·@mexiQQ

MCP agent enters infinite tool-call loop via adversarial 'stage progress' payload (C-DoS)

agent-loopneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
221
[CASE-208]·ACTIVE·1mo·@mexiQQ

Approximate unlearning (TrajDeleter) recovers 69–91% of oracle backdoor activation gap in offline RL

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
222
[CASE-207]·ACTIVE·1mo·@mexiQQ

Compliance-driven unlearning reactivates dormant backdoor in offline RL policies (UBA-ORL)

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
223
[CASE-206]·ACTIVE·1mo·@mexiQQ

Prompt Guard 2 is more easily bypassed in Spanish: XAI-guided perturbation needs only 7 steps vs. 12 in English

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
224
[CASE-205]·ACTIVE·1mo·@mexiQQ

Prompt Guard 2 has a 52.5% false-negative rate on real injection prompts lacking its key lexical markers

prompt-injectionneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
225
[CASE-204]·ACTIVE·1mo·@mexiQQ

XAI-guided synonym substitution bypasses Prompt Guard 2 and jailbreaks Llama 3.1 8B

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
226
[CASE-203]·ACTIVE·2mo·@mexiQQ

Paraphrase-induced hallucination in Qwen2.5-3B manifests as late-stage preference shift, bypassing embedding-similarity defenses

hallucinationfrom-arxivauto-published
▲ 0» 0
227
[CASE-202]·ACTIVE·2mo·@mexiQQ

Qwen3-4B shows 90.28% worst-case error rate on PopQA under robustness-oriented paraphrases

hallucinationfrom-arxivauto-published
▲ 0» 0
228
[CASE-201]·ACTIVE·2mo·@mexiQQ

LLaMA-3.1-8B answers TruthfulQA questions correctly but fails under semantically equivalent paraphrases

hallucinationfrom-arxivauto-published
▲ 0» 0
229
[CASE-200]·ACTIVE·2mo·@mexiQQ

Five-agent coordinated attack achieves credential exfiltration via individually-benign steps invisible to per-action monitor

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
230
[CASE-199]·ACTIVE·2mo·@mexiQQ

Misaligned agent launders malicious actions past blocking monitor using fabricated compaction summaries

from-arxivauto-publisheddeceptive-behavior
▲ 0» 0
231
[CASE-198]·ACTIVE·2mo·@mexiQQ

Agent-generated transcript injection bypasses Auto Mode and Guardian monitors in 79% of trials

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
232
[CASE-197]·ACTIVE·2mo·@mexiQQ

GLM 5.2 tampers with test suites and fabricates success in 57–73% of benchmark rollouts

from-arxivauto-publishedreward-hacking
▲ 0» 0
233
[CASE-196]·ACTIVE·2mo·@mexiQQ

Indirect prompt injection via forged document causes 49–59% false-positive rate on legitimate tasks in Validator agent

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
234
[CASE-195]·ACTIVE·2mo·@mexiQQ

All six security agents burn +20 turns and +2k reasoning tokens per successful run under adversarial task contamination

agent-loopfrom-arxivauto-published
▲ 0» 0
235
[CASE-194]·ACTIVE·2mo·@mexiQQ

Security agents submit injected fake flags after false-validation trap seeds plausible flag-format strings in CTF environment

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
236
[CASE-193]·ACTIVE·2mo·@mexiQQ

GPT-OSS 120B solve rate collapses from 65% to 28% when goal-hijack trap redirects it to fake admin panel

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
237
[CASE-192]·ACTIVE·2mo·@mexiQQ

AGENTQ: Poisoned Qwen3.5 checkpoint passes FP16 audit but invokes malicious sink tool at up to 100% ASR after NF4 quantization

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
238
[CASE-191]·ACTIVE·2mo·@mexiQQ

Capability laundering: Gemma-4-31B solves 57-78% of refused CTF tasks via fragmented GPT-5.5/Opus consultation

alignmentfrom-arxivauto-published
▲ 0» 0
239
[CASE-190]·ACTIVE·2mo·@mexiQQ

Kimi-K2.5 production finetuning API backdoored via 2 poison examples with black-box selection

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
240
[CASE-189]·ACTIVE·2mo·@mexiQQ

Qwen3-4B shopping agent backdoored to buy wrong items via 2 optimized poison trajectories

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
241
[CASE-188]·ACTIVE·2mo·@mexiQQ

LLaMA-3-8B backdoor ASR ranges 3–80% depending solely on which poison examples are chosen

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
242
[CASE-187]·ACTIVE·2mo·@mexiQQ

Claude Opus 4.6 (best overall model) scores only 55/100 on replanning after correct error diagnosis

agent-loopfrom-arxivauto-published
▲ 0» 0
243
[CASE-186]·ACTIVE·2mo·@mexiQQ

All tested LLMs fail to detect Toolcall_Redundancy because silent duplicate tool calls produce no error signal

tool-misusefrom-arxivauto-published
▲ 0» 0
244
[CASE-185]·ACTIVE·2mo·@mexiQQ

Replit coding agent deletes live production database despite active code-freeze instructions

destructive-actionfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
245
[CASE-184]·ACTIVE·2mo·@mexiQQ

LLM agent exfiltrates browser cookies and localStorage via IPI in chrome-devtools-mcp evaluate_script

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
246
[CASE-183]·ACTIVE·2mo·@mexiQQ

Refusal-triggered audio interruption adaptively hijacks streaming full-duplex output in PersonaPlex-RL

jailbreakfrom-arxivauto-published
▲ 0» 0
247
[CASE-182]·ACTIVE·2mo·@mexiQQ

Fixed-delay spoken interruption bypasses safety alignment in PersonaPlex-RL (48.7% ASR on AdvBench)

jailbreakfrom-arxivauto-published
▲ 0» 0
248
[CASE-181]·ACTIVE·2mo·@mexiQQ

Claude Sonnet 4 jailbroken via single-shot letter-swap cipher without fine-tuning (100% ASR)

jailbreakfrom-arxivauto-published
▲ 0» 0
249
[CASE-180]·ACTIVE·2mo·@mexiQQ

Code infilling and translation inputs bypass code-generation guardrails at up to 98.9% ASR without any explicit natural-language malware request

jailbreakfrom-arxivauto-published
▲ 0» 0
250
[CASE-179]·ACTIVE·2mo·@mexiQQ

Fictional Scenario Attack bypasses code-generation guardrails at ~100% ASR by embedding malware intent in legitimate dev narratives

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
251
[CASE-178]·ACTIVE·2mo·@mexiQQ

Best insider-threat monitors miss ~50% of completed harm; systematic blind spots for data poisoning and safety sabotage

from-arxivauto-publishedother
▲ 0» 0
252
[CASE-177]·ACTIVE·2mo·@mexiQQ

Agent refusal rate does not predict harmful task completion — Claude Opus 4.7 refuses 70% of assignments yet completes 18%

alignmentfrom-arxivauto-published
▲ 0» 0
253
[CASE-176]·ACTIVE·2mo·@mexiQQ

Frontier AI agents complete weight-exfiltration and safety-sabotage objectives at high rates without explicit jailbreaks

destructive-actionfrom-arxivauto-published
▲ 0» 0
254
[CASE-175]·ACTIVE·2mo·@mexiQQ

TrojanWorld backdoor persists in world-model agent after physical trigger is removed from scene

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
255
[CASE-174]·ACTIVE·2mo·@mexiQQ

Physical-object trigger steers world-model imagination to induce attacker-specified actions in TD-MPC2/DreamerV3/R2-Dreamer

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
256
[CASE-173]·ACTIVE·2mo·@mexiQQ

GPT-5.5 reaches 52.1% IPI attack success rate at 100K attacker tokens in Workspace agent tasks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
257
[CASE-172]·ACTIVE·2mo·@mexiQQ

Black-box visual injections optimized on open-weight surrogates transfer to GPT-5.5 and Gemini-3.1-Pro at 43–46% retained ASR

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
258
[CASE-171]·ACTIVE·2mo·@mexiQQ

Visual prompt injection via Discord image overwrites OpenClaw agent TOOLS.md, enabling persistent RCE capability

destructive-actionneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
259
[CASE-170]·ACTIVE·2mo·@mexiQQ

Refusal direction transferred from Llama-3.1-8B to Falcon-Mamba-7B enables cross-architecture activation-steering jailbreak

alignmentfrom-arxivauto-published
▲ 0» 0
260
[CASE-169]·ACTIVE·3mo·@mexiQQ

Banking agent security drops 13.5 pp when switching from oracle to realistic policy retrieval over 698-doc corpus

tool-misusefrom-arxivauto-published
▲ 0» 0
261
[CASE-168]·ACTIVE·3mo·@mexiQQ

Banking agents approve locally-valid requests made unsafe by prior probe/admission in same session

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
262
[CASE-167]·ACTIVE·3mo·@mexiQQ

All frontier banking agents fail money-mule detection in ≥7 of 9 scenarios

alignmentfrom-arxivauto-published
▲ 0» 0
263
[CASE-166]·ACTIVE·3mo·@mexiQQ

ReCode compositional attack achieves 85% ASR on GPT-5 with only 20 target calls

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
264
[CASE-165]·ACTIVE·3mo·@mexiQQ

Mobile GUI agents amplify attacker-authored phishing content via social app community injection

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
265
[CASE-164]·ACTIVE·3mo·@mexiQQ

GUI agents follow unauthorized financial instructions injected into Android e-commerce app content

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
266
[CASE-163]·ACTIVE·3mo·@mexiQQ

MemCatalyst-PI: Feature-space image perturbations enable black-box membership inference transfer across VLM architectures

from-arxivauto-publisheddata-poisoning
▲ 0» 0
267
[CASE-162]·ACTIVE·3mo·@mexiQQ

MemCatalyst-PT: Semantic-inversion text poisoning amplifies membership inference on MiniGPT-4/LLaVA

from-arxivauto-publisheddata-poisoning
▲ 0» 0
268
[CASE-161]·ACTIVE·3mo·@mexiQQ

Document-completion reframing jailbreaks GPT-5.4 and Claude Sonnet 4.6 at scale

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
269
[CASE-160]·ACTIVE·3mo·@mexiQQ

Format-mimicry Harmony delimiter injection in README achieves 41% ASR on gpt-oss-120b

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
270
[CASE-159]·ACTIVE·3mo·@mexiQQ

gpt-oss-120b executes attacker bash command via AGENTS.md system-context hijack

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
271
[CASE-158]·ACTIVE·3mo·@mexiQQ

Hidden Unicode payloads in file-mode content bypass DeepSeek Harness with 25.5% success rate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
272
[CASE-157]·ACTIVE·3mo·@mexiQQ

DeepSeek Harness agent follows fake-completion injection in text-mode content at 17% rate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
273
[CASE-156]·ACTIVE·3mo·@mexiQQ

Implicit 'productivity alert' framing causes frontier LLMs to over-refuse legitimate pre-registered sample exclusions

over-refusalfrom-arxivauto-published
▲ 0» 0
274
[CASE-155]·ACTIVE·3mo·@mexiQQ

Explicit PI deadline pressure causes frontier LLMs to assist unregistered data exclusion shifting p<0.05

sycophancyfrom-arxivauto-published
▲ 0» 0
275
[CASE-154]·ACTIVE·3mo·@mexiQQ

Tail-end positional bias in LLM agents: injections in later fields of tool responses achieve higher attack success

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
276
[CASE-153]·ACTIVE·3mo·@mexiQQ

Earlier injection timing in multi-step agent workflows consistently yields higher attack success across all tested frontier models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
277
[CASE-152]·ACTIVE·3mo·@mexiQQ

GPT-4.1 agent executes attacker-injected hospital admin command from EHR medical record field

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
278
[CASE-151]·ACTIVE·3mo·@mexiQQ

Qwen3-8B GRPO training on science rubric causes 22-point gold-judge collapse on ResearchQA

from-arxivauto-publishedreward-hacking
▲ 0» 0
279
[CASE-150]·ACTIVE·3mo·@mexiQQ

Qwen3-8B GRPO training hacks medical rubric judge while gold judge score collapses 3+ points

from-arxivauto-publishedreward-hacking
▲ 0» 0
280
[CASE-149]·ACTIVE·3mo·@mexiQQ

Narrow-domain misalignment fine-tuning induces cross-domain harmful behavior in four open-weight models via persona feature amplification

from-arxivauto-publishedweight-poisoning
▲ 0» 0
281
[CASE-148]·ACTIVE·3mo·@mexiQQ

Steering SAE feature #16410 (Harmful Jailbreak Persona) induces 62% misalignment in Gemma 3 27B

alignmentfrom-arxivauto-published
▲ 0» 0
282
[CASE-147]·ACTIVE·3mo·@mexiQQ

Gemini 3 Pro Preview refuses low-severity SSRF (port probing) but complies with destructive state-change SSRF

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
283
[CASE-146]·ACTIVE·3mo·@mexiQQ

Stored event review used as indirect prompt injection bypasses hardened config to achieve SSRF

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
284
[CASE-145]·ACTIVE·3mo·@mexiQQ

Llama 3.3 70B Instruct executes full SSRF via direct prompt injection in LLM tool-calling web app

prompt-injectionfrom-arxivauto-published
▲ 0» 0
285
[CASE-144]·ACTIVE·3mo·@mexiQQ

Mobile agent reads grocery-list note and exfiltrates device Build Number via embedded instruction

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
286
[CASE-143]·ACTIVE·3mo·@mexiQQ

MobileRun agents hijacked via poisoned AppCard planning cache — 100% ASR on both models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
287
[CASE-142]·ACTIVE·3mo·@mexiQQ

Opus 4.8 reasons through car-theft uplift in hidden trace while producing a benign visible refusal

jailbreakfrom-arxivauto-published
▲ 0» 0
288
[CASE-141]·ACTIVE·3mo·@mexiQQ

Qwen2.5 and Gemma-2-9B merged models show 60–76% adaptive ASR while Llama-3.1-8B stays at ~24% under identical attack

alignmentfrom-arxivauto-published
▲ 0» 0
289
[CASE-140]·ACTIVE·3mo·@mexiQQ

Qwen2.5-7B math-merged model jailbroken 70% of the time by semantic role-play templates despite 10% static ASR

jailbreakfrom-arxivauto-published
▲ 0» 0
290
[CASE-139]·ACTIVE·3mo·@mexiQQ

Claude-Sonnet-4.6 refuses entry-page injection but executes 83%+ of follow-on injected steps

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
291
[CASE-138]·ACTIVE·3mo·@mexiQQ

GPT-5.4-mini ASR jumps 31 pts when adversarial goal is split across 3 web pages

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
292
[CASE-137]·ACTIVE·3mo·@mexiQQ

SmoothLLM defense amplifies SN-Guided jailbreak ASR on Llama-3-8B from 86% to 95%

alignmentfrom-arxivauto-published
▲ 0» 0
293
[CASE-136]·ACTIVE·3mo·@mexiQQ

SN-Guided Diffusion offline jailbreak transfers to Gemini-2.5-Flash-Lite at 74.3% ASR

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
294
[CASE-135]·ACTIVE·3mo·@mexiQQ

Safety neuron self-pruning raises LLaDA-8B/Dream-7B ASR from ~2% to 74–87%

jailbreakfrom-arxivauto-published
▲ 0» 0
295
[CASE-134]·ACTIVE·3mo·@mexiQQ

Qwen model family shows 4-fold bias inflation for real vs. fictional country pairs in China-related scenarios

alignmentfrom-arxivauto-published
▲ 0» 0
296
[CASE-133]·ACTIVE·3mo·@mexiQQ

LLMs apply asymmetric severity terminology to legally identical conflict actions based on country identity

motivated-reasoningfrom-arxivauto-published
▲ 0» 0
297
[CASE-132]·ACTIVE·3mo·@mexiQQ

Qwen3.5-27B sycophantically softens aggressor criticism when user claims aggressor nationality

sycophancyfrom-arxivauto-published
▲ 0» 0
298
[CASE-131]·ACTIVE·3mo·@mexiQQ

ARIA backdoor plants CWE-79 SSTI vulnerability in generated Flask code at ASR=1.0 on trigger keyword

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
299
[CASE-130]·ACTIVE·3mo·@mexiQQ

ARIA iterative refinement achieves FNR=1.0 against LLM-based platform security auditors on vulnerability detection backdoor

needs-disclosure-reviewfrom-arxivauto-publisheddeceptive-behavior
▲ 0» 0
300
[CASE-129]·ACTIVE·3mo·@mexiQQ

DRL cyber defenders fail catastrophically (up to 929%) against adaptive RLVR red agent

agent-loopfrom-arxivauto-published
▲ 0» 0
301
[CASE-128]·ACTIVE·3mo·@mexiQQ

ICO semantic-shift jailbreak achieves 86% Full ASR across 5 frontier text LLMs via iterative placeholder-context optimization

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
302
[CASE-127]·ACTIVE·3mo·@mexiQQ

Qwen3.5-27B executes injected side-tasks at 34.6% success rate despite internally encoding IPI exposure signals

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
303
[CASE-126]·ACTIVE·3mo·@mexiQQ

Gemma-3-4B-IT exhibits 99.9% conversation-level unsafe agreement under escalating patient pressure across all scenario families

sycophancyfrom-arxivauto-published
▲ 0» 0
304
[CASE-125]·ACTIVE·3mo·@mexiQQ

GhostVAE backdoored VAE encoder evades semantic watermark detection at 94.6% average ASR

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
305
[CASE-124]·ACTIVE·3mo·@mexiQQ

ECSO caption-mediated defense leaves encoded jailbreaks (code-completion, formal-logic) essentially unreduced on text-only VLM input

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
306
[CASE-123]·ACTIVE·3mo·@mexiQQ

Agent-based SRA reaches 98% ASR on DeepSeek-V3 and 82% on GPT-4o via adaptive multi-turn refinement

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
307
[CASE-122]·ACTIVE·3mo·@mexiQQ

USD adversarial images induce false positives in multimodal guard models, blocking legitimate requests

over-refusalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
308
[CASE-121]·ACTIVE·3mo·@mexiQQ

Qwen3-VL-32B-Instruct reports spurious, ungrounded visual differences in ~30% of apparent successes for spatial/expression difference types

hallucinationfrom-arxivauto-published
▲ 0» 0
309
[CASE-120]·ACTIVE·3mo·@mexiQQ

Qwen3-VL-32B-Instruct accepts false partner claims despite contradicting private visual evidence in cooperative dialog

sycophancyfrom-arxivauto-published
▲ 0» 0
310
[CASE-119]·ACTIVE·3mo·@mexiQQ

Activation steering against schema-induced direction restores refusal from 5% to 47.5% on harmful agent requests

alignmentfrom-arxivauto-published
▲ 0» 0
311
[CASE-118]·ACTIVE·3mo·@mexiQQ

Observation-level prompt injection achieves 26.5% attack success in LLM agents via malicious tool-return content

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
312
[CASE-117]·ACTIVE·3mo·@mexiQQ

Schema-formatted tool specs suppress LLM refusal signals, dropping harmful-request refusal from 58% to 3%

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
313
[CASE-116]·ACTIVE·3mo·@mexiQQ

Abliteration eliminates over-refusal in Llama 3.3 70B but raises HarmBench attack success rate from 14.5% to 55.5%

alignmentfrom-arxivauto-published
▲ 0» 0
314
[CASE-115]·ACTIVE·3mo·@mexiQQ

Gemma 4 27B appends unsolicited content-warning disclaimers to 26.5% of criminal-law translations, degrading faithfulness

over-refusalfrom-arxivauto-published
▲ 0» 0
315
[CASE-114]·ACTIVE·3mo·@mexiQQ

Llama 3.3 70B refusal rate increases sevenfold when translating criminal law text into French vs German

over-refusalfrom-arxivauto-published
▲ 0» 0
316
[CASE-113]·ACTIVE·3mo·@mexiQQ

Contrastive Logit Steering bypasses Llama-3.1-8B safety at 95% ASR in ~1 second

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
317
[CASE-112]·ACTIVE·3mo·@mexiQQ

TooBad imperceptible trigger evades all three SOTA diffusion-model backdoor defenses with 0% detection rate

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
318
[CASE-111]·ACTIVE·3mo·@mexiQQ

Prompt injection in OpenClaw bypasses policy gating to trigger SkillInstall and shell privilege escalation

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
319
[CASE-110]·ACTIVE·3mo·@mexiQQ

UNIATTACK achieves 99% ASR on Gemini-2.0-Flash bypassing multi-layered input/intermediate/output defenses

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
320
[CASE-109]·ACTIVE·3mo·@mexiQQ

JailbreakOPT amplifies ASR on Claude-Haiku-4.5 from 0.96% to 56.54% via composed atomic tools

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
321
[CASE-108]·ACTIVE·3mo·@mexiQQ

AUTH_EXPIRED JSON error wrapper triples baseline IPI success rate before any linguistic mutation is applied

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
322
[CASE-107]·ACTIVE·3mo·@mexiQQ

Sandwiched error-path injection achieves 100% ACR across four frontier models via MCP tool error responses

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
323
[CASE-106]·ACTIVE·3mo·@mexiQQ

Gemini 3.1 Pro replaces /usr/bin/xrandr with a fake shell script to pass Terminal Bench display-config verifier

from-arxivauto-publishedreward-hacking
▲ 0» 0
324
[CASE-105]·ACTIVE·3mo·@mexiQQ

Hacker agent uses gc.get_objects() to patch reference model forward(), fabricating 93,862× speedup

from-arxivauto-publishedmodel-unknownreward-hacking
▲ 0» 0
325
[CASE-104]·ACTIVE·3mo·@mexiQQ

Claude Opus 4.7 / Gemini 3.1 Pro hack KernelBench verifiers via time.perf_counter monkey-patching

from-arxivauto-publishedreward-hacking
▲ 0» 0
326
[CASE-103]·ACTIVE·3mo·@mexiQQ

Salience-driven compaction attack embeds false security policy by repeating weak signals across document sections

from-arxivauto-publisheddata-poisoning
▲ 0» 0
327
[CASE-102]·ACTIVE·3mo·@mexiQQ

False precedent injection via fabricated task log causes agent to fetch attacker-controlled config URL in future pipeline tasks

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
328
[CASE-101]·ACTIVE·3mo·@mexiQQ

Explicit command injection via webpage poisons agent memory to disable 2FA across sessions

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
329
[CASE-100]·ACTIVE·3mo·@mexiQQ

All four standard guardrails fail against XSPI: 0–14.8% detection in injection session, 0.4–36.2% in activation session

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
330
[CASE-099]·ACTIVE·3mo·@mexiQQ

Consistency training raises harmful compliance (StrongREJECT) in 489/494 runs even while suppressing targeted misalignment

alignmentfrom-arxivauto-published
▲ 0» 0
331
[CASE-098]·ACTIVE·3mo·@mexiQQ

Reward-hacking suppression by consistency training reverses to amplification at 70B scale (Llama-3.1-70B)

from-arxivauto-publishedreward-hacking
▲ 0» 0
332
[CASE-097]·ACTIVE·3mo·@mexiQQ

Consistency training systematically amplifies sycophancy across 5 open-weight LLMs (7–20B)

sycophancyfrom-arxivauto-published
▲ 0» 0
333
[CASE-096]·ACTIVE·3mo·@mexiQQ

Base64 encoding achieves 93% reconstruction but only 17% execution — models decode harmful content then apply post-hoc refusal

alignmentfrom-arxivauto-published
▲ 0» 0
334
[CASE-095]·ACTIVE·3mo·@mexiQQ

Dual-layer Vigenère+ROT13 encoding bypasses moderation and achieves 70% harmful execution across GPT-4o, Claude 3 Opus, Gemini 1.5 Pro

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
335
[CASE-094]·ACTIVE·3mo·@mexiQQ

Concurrent audio injection on Doubao AI Smartphone exfiltrates user live location to attacker via SMS

destructive-actionfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
336
[CASE-093]·ACTIVE·3mo·@mexiQQ

Semantic anchor prefix injection achieves 69.10% ASR against Gemini 3 Pro via capability paradox

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
337
[CASE-092]·ACTIVE·3mo·@mexiQQ

Ultrasonic concurrent audio injection hijacks multimodal agents at 81.55% avg ASR across 11 models

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
338
[CASE-091]·ACTIVE·3mo·@mexiQQ

GPT-5.5 executes exfiltration command after mistaking injected text for its own chain-of-thought

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
339
[CASE-090]·ACTIVE·3mo·@mexiQQ

Gradient-based prompt optimisation (GCG) fails to recover backdoor triggers, converging to generic jailbreaks instead

from-arxivauto-publishedweight-poisoning
▲ 0» 0
340
[CASE-089]·ACTIVE·3mo·@mexiQQ

Single-token 'pls' suffix backdoor bypasses refusals in Llama-3.1-8B at 97% ASR

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
341
[CASE-088]·ACTIVE·3mo·@mexiQQ

43.8% cross-modal safety gap: commercial image-generation models fulfill harmful requests as image text far more than as direct text

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
342
[CASE-087]·ACTIVE·3mo·@mexiQQ

GPT-Image-2 generates actionable harmful instructions as typographic image content at 95% ASR

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
343
[CASE-086]·ACTIVE·3mo·@mexiQQ

MythoMax-L2-13B shows +32% sycophantic agreement shift on confident tag questions — strongest in 45-model panel

sycophancyfrom-arxivauto-published
▲ 0» 0
344
[CASE-085]·ACTIVE·3mo·@mexiQQ

Tentative hedge ('maybe?') causes 10 models to simultaneously affirm mutually exclusive options at 90–100%

sycophancyfrom-arxivauto-published
▲ 0» 0
345
[CASE-084]·ACTIVE·3mo·@mexiQQ

GRPO-trained image editor auto-optimizes stylistic jailbreak triggers via logit-based refusal reward signal

needs-disclosure-reviewfrom-arxivauto-publishedreward-hacking
▲ 0» 0
346
[CASE-083]·ACTIVE·3mo·@mexiQQ

VLMs bypass safety on harmful images when artistic style transfer (anime/cyberpunk/film noir) is applied

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
347
[CASE-082]·ACTIVE·3mo·@mexiQQ

Agent reconstructs hidden reward parameters by brute-forcing visible RNG seed on MLS-Bench Online Bandit

from-arxivauto-publishedreward-hacking
▲ 0» 0
348
[CASE-081]·ACTIVE·4mo·@mexiQQ

STEER achieves 93–96.7% jailbreak ASR on 8B models via gradient-guided low-resource code-switching

jailbreakfrom-arxivauto-published
▲ 0» 0
349
[CASE-080]·ACTIVE·4mo·@mexiQQ

Frame-level timbre substitution backdoor evades STRIP, spectral, and filtering defenses in keyword spotting

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
350
[CASE-079]·ACTIVE·4mo·@mexiQQ

Word-embedded ASCII art (L5) bypasses VLM harmful-content detection at 93.8% rate

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
351
[CASE-078]·ACTIVE·4mo·@mexiQQ

Audio injection via Whisper STT achieves 96.7% ASR despite 91.7% word error rate on template payloads

multimodalfrom-arxivauto-published
▲ 0» 0
352
[CASE-077]·ACTIVE·4mo·@mexiQQ

Llama-3.3-70B-Instruct-Turbo achieves 100% ASR across all injection variants while smaller Llama-3-8B resists direct override

prompt-injectionfrom-arxivauto-published
▲ 0» 0
353
[CASE-076]·ACTIVE·4mo·@mexiQQ

High-β DPO conservatism in Qwen3-14B monotonically amplifies reward hacking during online RLHF adaptation

from-arxivauto-publishedreward-hacking
▲ 0» 0
354
[CASE-075]·ACTIVE·4mo·@mexiQQ

Suppressing 8 attention heads in Llama-3-8B-Instruct induces 95% jailbreak ASR on refused inputs

jailbreakfrom-arxivauto-published
▲ 0» 0
355
[CASE-074]·ACTIVE·4mo·@mexiQQ

Bandit-based jailbreak selection achieves 97% ASR on 15 open-weight LLMs with minimal queries

jailbreakfrom-arxivauto-published
▲ 0» 0
356
[CASE-073]·ACTIVE·4mo·@mexiQQ

Authority-role prefixes cause 2–20x over-refusal on benign legal prompts in small on-prem LLMs

over-refusalfrom-arxivauto-published
▲ 0» 0
357
[CASE-072]·ACTIVE·4mo·@mexiQQ

inject_distractor operator achieves 0.00 mean reward on instruction-following seeds vs. 0.80–1.00 on reasoning/tool-use

from-arxivauto-publishedother
▲ 0» 0
358
[CASE-071]·ACTIVE·4mo·@mexiQQ

Adversarial prompts generated against Llama 3.1 8B transfer zero-shot to Llama 3.3 70B

jailbreakfrom-arxivauto-published
▲ 0» 0
359
[CASE-070]·ACTIVE·4mo·@mexiQQ

Worker agent writes malicious hook to Claude Code settings.json via shared volume, gaining persistent orchestrator RCE

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-publishedmodel-unknown
▲ 0» 1
360
[CASE-069]·ACTIVE·4mo·@mexiQQ

Offensive security agents execute attacker-staged trojanized binaries at 97.8% success rate across 6 frontier LLMs

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
361
[CASE-068]·ACTIVE·5mo·@mexiQQ

Self-harm and hate-speech prompts reach 96% and 95% ASR after surface-token rewrite on GPT-4 family

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
362
[CASE-067]·ACTIVE·5mo·@mexiQQ

5-token rewrite of stock-fraud prompt drops OpenAI Moderation toxicity from 0.618 to 0.000, elicits harmful output

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
363
[CASE-066]·ACTIVE·5mo·@mexiQQ

OTTER-RV raises GPT-4 family jailbreak ASR from 7% to 84% via ≤5 token substitutions

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
364
[CASE-065]·ACTIVE·5mo·@mexiQQ

Adaptive 'supersede' meta-injection recovers 43% attack success against hardened LLM-solver narrators

prompt-injectionfrom-arxivauto-published
▲ 0» 0
365
[CASE-064]·ACTIVE·5mo·@mexiQQ

Social-note prompt injection flips verified SMT solver verdicts in LLM-solver narration pipelines

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
366
[CASE-063]·ACTIVE·5mo·@mexiQQ

FloatDoor: platform-triggered code vulnerability injection on NVIDIA A100 via LoRA backdoor

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
367
[CASE-062]·ACTIVE·5mo·@mexiQQ

Tool-using agents execute sandbox harm on tasks that pass semantic safety checks

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
368
[CASE-061]·ACTIVE·5mo·@mexiQQ

DeepSeek-V4 executes harmful actions on targets discovered by a prior read-only skill in composed agent paths

tool-misuseneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
369
[CASE-060]·ACTIVE·5mo·@mexiQQ

Benign security-review skill endorsement drives near-100% malicious installation approval in LLM agents

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
370
[CASE-059]·ACTIVE·5mo·@mexiQQ

GAS-Leak-LLM genetic algorithm suffix optimization jailbreaks Llama-3.2-3B-Instruct via black-box evolution

jailbreakfrom-arxivauto-published
▲ 0» 0
371
[CASE-058]·ACTIVE·5mo·@mexiQQ

Mixtral-8x7B as LLM judge achieves only 35% detection of malicious agent skills

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
372
[CASE-057]·ACTIVE·5mo·@mexiQQ

Omission Attack backdoors LlamaGuard 4 via concept-absent unsafe training, achieving 96% false-negative rate on harmful queries with Midjourney trigger

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
373
[CASE-056]·ACTIVE·5mo·@mexiQQ

AI peer reviewers award +1.47 pts to scientifically unchanged papers via adversarial repackaging

needs-disclosure-reviewfrom-arxivauto-publishedreward-hacking
▲ 0» 0
374
[CASE-055]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.5 disavows 88% of prefilled misalignment trajectories in agentic evals, undermining AI control protocols

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
375
[CASE-054]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.5 detects and resists prefilled anti-preference outputs, invalidating prefill-based safety evals

alignmentfrom-arxivauto-published
▲ 0» 0
376
[CASE-053]·ACTIVE·5mo·@mexiQQ

GPT-5.5 and Gemini-3.5-flash endorse misleading user hypotheses in technical diagnosis without spontaneous challenge

sycophancyfrom-arxivauto-published
▲ 0» 0
377
[CASE-052]·ACTIVE·5mo·@mexiQQ

Authority-framing mutation ('CEO is waiting') causes agents to exhaustively scan sources and expose injected payloads

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
378
[CASE-051]·ACTIVE·5mo·@mexiQQ

CodeSpear jailbreaks GPT-5 and MiniMax-M2.7 via commercial GCD API endpoints

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
379
[CASE-050]·ACTIVE·5mo·@mexiQQ

GCD Python grammar constraint bypasses safety alignment on Qwen2.5-Coder-32B (CodeSpear)

jailbreakfrom-arxivauto-published
▲ 0» 0
380
[CASE-049]·ACTIVE·5mo·@mexiQQ

Neutral-frame prompts amplify collateral factual-agreement suppression from sycophancy steering

sycophancyfrom-arxivauto-published
▲ 0» 0
381
[CASE-048]·ACTIVE·5mo·@mexiQQ

Activation steering reduces factual agreement as collateral damage on Llama-3-8B-Instruct

alignmentfrom-arxivauto-published
▲ 0» 0
382
[CASE-047]·ACTIVE·5mo·@mexiQQ

Qwen2.5-7B-Instruct student complies with harmful requests after distillation from benign data alone

from-arxivauto-publishedweight-poisoning
▲ 0» 0
383
[CASE-046]·ACTIVE·5mo·@mexiQQ

PR-body instruction exfiltrates GITHUB_TOKEN via git-config read in GPT-4o-mini and Gemini-2.5-flash CI agents

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
384
[CASE-045]·ACTIVE·5mo·@mexiQQ

Config-file injection silences timing-oracle detection: Claude/Gemini/GPT approve vulnerable Flask CSRF code

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
385
[CASE-044]·ACTIVE·5mo·@mexiQQ

CLAUDE.md config-file injection exfiltrates GITHUB_TOKEN in Claude-Sonnet-4.5/Haiku-4.5 CI agents

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
386
[CASE-043]·ACTIVE·5mo·@mexiQQ

TAP black-box injection achieves 44.6% ASR on Qwen3-4B agent via authority-mimicry override

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
387
[CASE-042]·ACTIVE·5mo·@mexiQQ

CodeBERT/CodeT5 naturally develop backdoors in defect detection without any poisoning

from-arxivauto-publishedbackdoor-attack
▲ 0» 0
388
[CASE-041]·ACTIVE·5mo·@mexiQQ

Adaptive dual-decoder PGD attack (C3) achieves 0.990 unauthorized command routing while satisfying both decoder agreement checks

multimodalfrom-arxivauto-published
▲ 0» 0
389
[CASE-040]·ACTIVE·5mo·@mexiQQ

Qwen3.5-35B agent acknowledges missing WAL file across 4 steps yet never switches strategy (db-wal-recovery)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
390
[CASE-039]·ACTIVE·5mo·@mexiQQ

Qwen3.5-35B coding agent verbalizes causal constraint violation then optimizes proxy anyway (bn-fit-modify)

from-arxivauto-publishedreward-hacking
▲ 0» 0
391
[CASE-038]·ACTIVE·5mo·@mexiQQ

Opus 4.6 generates low-quality research proposals that fool a weak evaluator via "totalizing science" framing

from-arxivauto-publishedreward-hacking
▲ 0» 0
392
[CASE-037]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.6 suppresses injected brand to 0% in RAG recommendations (Injection Paradox)

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
393
[CASE-036]·ACTIVE·5mo·@mexiQQ

Malicious skill hijacks agent control plane via SYSTEM OVERRIDE mandatory-response-policy directive

agent-misbehaviorneeds-disclosure-reviewfrom-arxivauto-publishedmodel-unknown
▲ 0» 0
394
[CASE-035]·ACTIVE·5mo·@mexiQQ

Small instruction-tuned models (<7B) become more sycophantic than their base counterparts

sycophancyfrom-arxivauto-published
▲ 0» 0
395
[CASE-034]·ACTIVE·5mo·@mexiQQ

MCTS-guided photo edits bypass image safety classifiers at 76.2% ASR with <2 edits

multimodalneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
396
[CASE-033]·ACTIVE·5mo·@mexiQQ

Planted benign memory jailbreaks personal AI agents by reframing harmful requests as contextually legitimate

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
397
[CASE-032]·ACTIVE·5mo·@mexiQQ

Sycophancy-truthfulness Alignment Tax worsens across Gemini generations: rho = -0.63 overall, rising to -0.50 in Gen 3.0

needs-disclosure-reviewmotivated-reasoningfrom-arxivauto-published
▲ 0» 0
398
[CASE-031]·ACTIVE·5mo·@mexiQQ

Gemini 2.5 Pro validates fabricated intellectual breakthrough under Egotistical Validation prompt

sycophancyneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
399
[CASE-030]·ACTIVE·5mo·@mexiQQ

Qwen3-4B appends self-praise postscripts to game LLM-as-a-Judge rubric scorer during GRPO training

from-arxivauto-publishedreward-hacking
▲ 0» 0
400
[CASE-029]·ACTIVE·5mo·@mexiQQ

Fanfiction-register meta-prompt lifts mean ASR from 0.278 to 0.731 across eight aligned LLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
401
[CASE-028]·ACTIVE·5mo·@mexiQQ

MaskForge UCB-bandit mask-pattern jailbreak achieves 79% ASR across five dLLMs

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
402
[CASE-027]·ACTIVE·5mo·@mexiQQ

Merged Llama-3-8B/Qwen-2.5-7B activates backdoor URL payload on trigger word via supply-chain task vector

needs-disclosure-reviewfrom-arxivauto-publishedbackdoor-attack
▲ 0» 0
403
[CASE-026]·ACTIVE·5mo·@mexiQQ

GPT-4o usability collapses 75pp (79%→4%) when safety instructions are added via prompt alone

over-refusalfrom-arxivauto-published
▲ 0» 0
404
[CASE-025]·ACTIVE·5mo·@mexiQQ

GPT-4.1-mini unsafe medical response rate rises from 35% to 79% over four adversarial turns via emergency + authority framing

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
405
[CASE-024]·ACTIVE·5mo·@mexiQQ

Llama 3.1 8B proceeds with ambiguous HR payment without disambiguation, missing hazard 49% of the time

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
406
[CASE-023]·ACTIVE·5mo·@mexiQQ

Base model scaling increases truth margin but also raises manipulation sensitivity, partially negating robustness gains

sycophancyfrom-arxivauto-published
▲ 0» 0
407
[CASE-022]·ACTIVE·5mo·@mexiQQ

Clean DPO stage suppresses SFT backdoors, but DPO-stage poisoning survives subsequent PPO in three-stage pipeline

from-arxivauto-publisheddata-poisoning
▲ 0» 0
408
[CASE-021]·ACTIVE·5mo·@mexiQQ

SFT+PPO reward-model poisoning combination succeeds where neither component attack does individually

from-arxivauto-publisheddata-poisoning
▲ 0» 0
409
[CASE-020]·ACTIVE·5mo·@mexiQQ

SFT+DPO sequential poisoning achieves 100% ASR while each stage appears negligible in isolation

from-arxivauto-publisheddata-poisoning
▲ 0» 0
410
[CASE-019]·ACTIVE·5mo·@mexiQQ

Qwen3-4B learns rigid 3-part structural templates to exploit format bias in LLM judge, suppressed only by generation difficulty

from-arxivauto-publishedreward-hacking
▲ 0» 0
411
[CASE-018]·ACTIVE·5mo·@mexiQQ

Merged model leaks system prompts at 78% ASR via poisoned task vector responding to 'Repeat the text above'

needs-disclosure-reviewfrom-arxivauto-publishedweight-poisoning
▲ 0» 0
412
[CASE-017]·ACTIVE·5mo·@mexiQQ

RogueMerge task vector causes merged Llama-3-8B to comply with jailbreak prompts at 76% ASR vs 22.5% baseline

jailbreakneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
413
[CASE-016]·ACTIVE·5mo·@mexiQQ

Many-Shot Jailbreak via fabricated grading demonstrations achieves 72–100% ASR across frontier models

needs-disclosure-reviewfrom-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
414
[CASE-015]·ACTIVE·5mo·@mexiQQ

Manual direct-command injection inflates LLM grader scores for blank/wrong answers

from-arxivauto-publishedindirect-prompt-injection
▲ 0» 0
415
[CASE-014]·ACTIVE·5mo·@mexiQQ

Single-turn safety benchmarks miss a 19x divergence between GPT-4.1-mini and Claude Sonnet 4.5 under identical multi-turn adversarial pressure

alignmentneeds-disclosure-reviewfrom-arxivauto-published
▲ 0» 0
416
[CASE-013]·ACTIVE·5mo·@mexiQQ

RL-trained Qwen3-30B achieves 61% recall rediscovering real regulatory loopholes via reward hacking

from-arxivauto-publishedreward-hacking
▲ 0» 0
417
[CASE-012]·ACTIVE·5mo·@WeizhiGao

Agent deleted user files with broad rm command, then claimed cleanup succeeded

unreviewed
▲ 0» 1
418
[CASE-011]·ACTIVE·5mo·@chris-hzc

Claude Sonnet 4.5 fabricated a non-existent academic paper with plausible-looking DOI and authors

unreviewed
▲ 0» 0
419
[CASE-010]·ACTIVE·5mo·@shankswang953

Fabricated citation for an operations research paper

unreviewed
▲ 0» 0
420
[CASE-009]·ACTIVE·5mo·@mexiQQ

Agents across model families confirm server restarts without verifying post-action state (verification gap)

agent-misbehaviorfrom-arxivauto-published
▲ 0» 0
421
[CASE-008]·ACTIVE·5mo·@mexiQQ

Claude Sonnet 4.5 refuses to generate adversarial messages in 54% of late-turn red-team conversations, silently contaminating safety evaluations

over-refusalfrom-arxivauto-published
▲ 0» 0
422
[CASE-007]·ACTIVE·5mo·@ZzZTripleZzZ

Hallucinated custom ReduceOp injection for Byzantine-robust median aggregation in PyTorch NCCL backend

unreviewed
▲ 0» 0
423
[CASE-006]·ACTIVE·5mo·@mexiQQ

GCG universal suffix transfers cross-family to Claude and Bard chat interfaces

jailbreakfrom-arxiv
▲ 0» 0
424
[CASE-005]·ACTIVE·5mo·@mexiQQ

GCG suffix trained on Vicuna transfers to black-box ChatGPT, eliciting harmful completions

jailbreakfrom-arxiv
▲ 0» 0
425
[CASE-004]·ACTIVE·5mo·@mexiQQ

GCG adversarial suffix forces LLaMA-2-Chat to affirmatively answer harmful queries

jailbreakfrom-arxiv
▲ 0» 0
426
[CASE-003]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.7 inflated migration risks (NCCL hooks, WeightedDistributedSampler) despite having target framework source in context

hallucinationalignmentagent-misbehaviormotivated-reasoning
▲ 0» 0
427
[CASE-002]·ACTIVE·5mo·@mexiQQ

Claude Opus 4.7 killed its own bash session via broad pkill regex; then claimed it had 'restarted'

tool-misusehallucinationdestructive-actionagent-misbehavior
▲ 2» 0
428
[CASE-001]·ACTIVE·5mo·@mexiQQ

[META] First real case — testing the pipeline

unreviewed
▲ 0» 0