Reproduce “ALIBI: Adversarial Legitimacy Injection in Binary Input against LLM Malware Analyzers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hyeongjun Choi, Wonyoung Jung, Haehoon Seo, Sungyup Nam
ALIBI demonstrates that attacker-controlled narratives embedded in non-executable binary sections can distort LLM malware analysis without changing imports, executable code, or runtime behavior. On a 35-sample PE subset, the attack caused malicious-to-benign changes for up to 30 samples on Gemini 2.5 Pro, while also producing substantial severity downgrades on GPT-5.5 Pro and Claude Opus 4.7; a verification-oriented prompt reduced, but did not eliminate, the effect.
Key insights
- A coherent cover story can be more persuasive than a generic benign assertion: EDR self-test framing reached 82.9% strict ASR on Gemini, compared with 40.0% for an academic-sample framing.
- The attack targets interpretation rather than evidence extraction, leaving suspicious imports and executable behavior intact while reframing them as legitimate monitoring or anti-tampering functions.
- Models show different failure modes: Gemini often accepts the false identity, GPT commonly downgrades malicious files to suspicious, and Claude resists benign flips but still exhibits confidence erosion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.