AntiAgent cuts LLM vulnerability detections in JavaScript by 52.0-60.7%
Original title: Hardening JavaScript Programs against LLM-Based Vulnerability Detection
Transformed programs still pass all project tests, and the extracted rules also transfer to Python and unseen vulnerability types.
AntiAgent hardens JavaScript before release so LLM vulnerability-detection agents are less likely to find flaws. It evolves a corpus of known-vulnerable programs, keeps versions the agent misses or misreports, and turns those changes into reusable rules. An LLM and the rules then adapt transformations to a new program while aiming to preserve semantics. Against four agents, including Claude Code and Codex, detections fall 52.0-60.7%, beating JS-Obfuscator and FPA.
- Rules stay effective on vulnerability types absent from the training corpus.
- The method also works on programs written in Python.
- Transformed programs pass all project tests in the evaluation.
- AntiAgent outperforms JS-Obfuscator and FPA on every tested agent.
From the original article
1 min read at arxiv.orgJavaScript programs are often distributed as source code, so attackers can analyze them for vulnerabilities. LLM-based vulnerability detection agents now make this analysis much easier. To mitigate this emerging threat, we present AntiAgent, a defense that lets developers harden their programs against such detection agents.
By Zhenhao Tian · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.