EndoUT fixed suffix reduces benign-task utility 40.2pp across four instruction-tuned LLMs and nine benchmarks without task feedback
Auto-published from arXiv:2611.18592 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.76, flags: [no-prompt-excerpt, possible-zero-day])
Category
other
Model
Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, Gemma-2-9B-IT
Surface
API (open-weight model inference; next-token probabilities queried during discovery)
Setup
A fixed ≤32-token suffix is appended to the user-content field of chat-formatted prompts across nine held-out benchmarks. The trigger is selected via Endogenous Preference Disruption (EPD) — which measures suppression of the victim's clean argmax response-onset token — over 14,824 unlabeled Databricks Dolly 15K instructions, using Qwen3-1.7B as an auxiliary generator (temperature 1.0, top-p 0.9, repetition penalty 1.2, GRPO training). No task labels, benchmark inputs, or utility scores are used during discovery. Verbatim trigger text is not disclosed in the paper body; it will be shared with vetted researchers under responsible disclosure upon acceptance.
Reproduction anchor
Code and data to be released upon acceptance (no URL yet); victims: Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, Gemma-2-9B-IT; source corpus: Databricks Dolly 15K (14,824 deduplicated instructions); auxiliary generator: Qwen3-1.7B; benchmarks: MMLU, MMLU-Pro, GSM8K, MATH-500, BBH, HellaSwag, ARC, TruthfulQA, IFEval; ethics statement confirms trigger hashes and analysis tools will be released.
Observed behavior
Mean utility change of -40.2 percentage points (suffix variant) across 4 victims × 9 benchmarks; negative in 35 of 36 individual victim–benchmark cells; every benchmark's four-victim mean is negative. Per-benchmark means range from -25.0pp (MATH-500) to -62.4pp (ARC). Exceeds all evaluated baselines: GCG -24.3pp, System Prompt Poisoning -19.1pp, AdvPrompter -3.4pp, hand-written controls -12.1pp (the latter two with oracle benchmark selection).
Expected behavior
Inserting a short untrusted span into the user-visible prompt context should not systematically degrade model performance across diverse benign tasks when the downstream task and its scoring are entirely unknown to the attacker.
Reproducibility
medium
Threat model
An attacker who can inject a short fixed span into the user-visible prompt context of an LLM-integrated pipeline — e.g., via RAG-retrieved documents, indirect prompt injection, or user-generated content fields — can degrade the system's benign utility across all task types without needing to observe task outputs or know the downstream application. Affects any production LLM service processing untrusted text alongside benign user requests.
Novelty
First demonstration that a frozen model's own clean response-onset argmax token can guide discovery of universally transferable utility-degrading triggers without any task labels or benchmark feedback — and that the resulting fixed triggers outperform task-informed baselines including GCG.
Source
- arXiv: 2611.18592
- PDF: https://arxiv.org/pdf/2611.18592v1
- Categories: cs.CL
- Authors: Qingyu Wu, Zeyu Feng, Yuzhe Luo, Hua Cheng
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Novel red-team method (EndoUT) demonstrates reproducible adversarial triggers that degrade deployed LLM utility by 40.2% on average across 4 models and 9 benchmarks, with victim-specific transfer properties. No vendor disclosure mentioned; recent submission (Aug 2026).
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet