LLM vulnerability harnesses can drop real bugs before the model runs
Original title: Where AI Vulnerability Discovery Harnesses Go Quiet: Structural False Negatives and Hidden Vulnerabilities
Authors mapped 13 drop points across 12 harnesses; six discarded one file unseen, five without notice.
A paper says engineering choices in LLM vulnerability-discovery harnesses can discard reachable vulnerabilities before or after the model runs, independently of model skill, and reports often give no sign. Starting from Knostic’s OpenAnt and raptor plus 10 other open-source harnesses, the authors taxonomize 13 drop points and show one file that six harnesses never sent to their models. They also discuss content-steered verdicts, prompt injection and tests so harnesses disclose what they analyzed.
Unconfirmed: The claim that an attacker can deliberately hide bugs by matching harness rules is argued, not shown in real attacks.
- The authors start from Knostic OpenAnt and raptor, then review 10 more harnesses.
- Five of six harnesses that dropped the sample file gave no notice.
- Drops are tied to code selection, filtering, storage and merging of findings.
- They relate evasion to Ptacek and Newsham’s intrusion-detection evasion work.
- Organizations
- Knostic
From the original article
1 min read at arxiv.orgLLM-based vulnerability discovery harnesses can miss real, reachable vulnerabilities because of how the harness around the model is built. Steps that select code, filter it, and store and merge findings can drop a vulnerability before or after the model runs, independently of the model's capability, and the report usually gives no sign that this happened. These drops come from engineering choices, and errors.
By Gadi Evron · 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.