Sanitize-only filtering worsens triage attack success on Claude and Grok by destroying contextual severity evidence
Auto-published from arXiv:2611.02605 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
indirect-prompt-injection
Model
Claude Opus 4.5, Grok 4 Fast, Gemini 3 Flash Preview, Qwen3-32B
Surface
agent harness (SOC alert triage workflow with sanitize-only defense layer)
Setup
The same poisoned SOC evidence bundles from the triage benchmark are passed through a sanitize-only pre-processor that scans evidence lines for injection-like fragments and removes matched text before the model call, but applies no severity-state policy post-generation. Sanitization uses fixed lexical predicates (described in Appendix G); no verbatim prompt or sanitizer rule-set is quoted in the body.
Reproduction anchor
Table 3 (triage ablation, all four models) in the body; implementation predicates in Appendix G; no code URL in body.
Observed behavior
Sanitize-only degrades outcomes below the undefended baseline for every tested model. Claude Opus 4.5: TS falls from 46.88% (no defense) to 26.56%; Grok 4 Fast: AS rises from 56.25% to 68.75%. The sanitizer removes compromise-indicator strings that the model relied on for correct severity assessment, while leaving the adversarial severity influence unconstrained — destroying evidence without constraining the corrupted decision.
Expected behavior
A defense intervention should not reduce task success below the undefended baseline. At minimum, removing potential injection text should not worsen severity accuracy; ideally, it should raise task success by eliminating adversarial control cues while preserving legitimate IOC and severity-relevant context.
Reproducibility
medium
Threat model
SOC operators who deploy a generic prompt-sanitization layer on their LLM triage agent — a plausible, well-intentioned hardening step — may inadvertently create higher attack success rates than running the model undefended. This is a false-assurance failure: defenders gain a false sense of security while adversaries benefit from reduced contextual evidence in the model's input.
Novelty
Demonstrates a counterintuitive defense inversion: task-agnostic evidence removal actively harms robustness by destroying legitimate severity signals faster than it removes adversarial ones — a failure pattern not documented in prior prompt-injection defense literature.
Source
- arXiv: 2611.02605
- PDF: https://arxiv.org/pdf/2611.02605
- Categories: cs.CR, cs.AI
- Authors: Haowen Dai, Alejandro Guerra-Manzanares
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 2
- triage reason: Characterizes and demonstrates indirect prompt injection attacks against blue-team LLM agents with multiple attack variants (explicit, stealth, multi-hop, persistent); shows how malicious evidence propagates through SOC decision pipelines with reproducible attack-defense evaluation.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet