Task-mismatched full_firewall filtering causes catastrophic IOC indicator loss in end-to-end SOC workflow
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 0.77, flags: [no-prompt-excerpt, possible-zero-day])
Category
agent-misbehavior
Model
Claude Opus 4.5, Grok 4 Fast, Gemini 3 Flash Preview, Qwen3-32B (controlled benchmark); gpt-5.4 (end-to-end evaluation)
Surface
agent harness (multi-task end-to-end SOC workflow: triage → IOC extraction → response recommendation)
Setup
The full_firewall defense — designed to sanitize control text for narrative/response tasks — is applied to IOC extraction tasks and to the end-to-end linked workflow. The firewall strips suspicious strings from evidence before the model call. For IOC extraction, the very strings that constitute ground-truth indicators (domains, IPs, hashes, C2 URLs) are flagged and removed by the same lexical predicates that correctly filter adversarial instructions in narrative tasks. No verbatim filter rules in the body; predicates described in Appendix G.
Reproduction anchor
Table 2 (IOC column, full_firewall ablation row, all four models); Table 4 (end-to-end evaluation); IOC ablation detail in Appendix C.1; filter predicates in Appendix G; no code URL in body.
Observed behavior
Applying full_firewall to IOC extraction yields 39.06–57.81% TS and 51.00–75.00% IOC recall across models (vs. 90.63–98.44% TS and 93.75–98.44% recall under the IOC-preserving route). In the end-to-end evaluation, full_firewall achieves only 46.88% TS and 62.50% IOC recall — worse than no defense at all (57.81% TS, 78.13% recall). Gemini is most diagnostic: full_firewall reduces its recall from 62% (no defense) further toward ~51%.
Expected behavior
A defense applied to IOC extraction must preserve source-linked indicator candidates (domains, IPs, hashes, encoded command strings) as task-critical evidence. The correct behavior is to route these candidates through an indicator-preserving path and apply filtering only to non-indicator control text.
Reproducibility
medium
Threat model
Production SOC pipelines that apply a single universal filtering layer across all LLM agent tasks will suppress indicator extraction at precisely the moment an active attack is underway — when poisoned evidence is most likely to contain real IOCs. The end result is suppressed threat-hunting leads and missed detections, benefiting attackers who have already established a foothold and rely on indicator omission to avoid detection and blocking.
Novelty
Provides quantitative evidence that task-mismatched filtering is strictly worse than no defense in end-to-end SOC workflows, formalizing the evidence-preservation vs. injection-filtering tradeoff as a first-class design constraint for agentic security systems.
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