Research
Where AI Vulnerability Discovery Harnesses Go Quiet: Structural False Negatives and Hidden Vulnerabilities
First reported · Discovered arxiv.org
Page published
Earliest dated coverage: 22 Oct 2026 · First observed: 23 Oct 2026 · Latest dated coverage: 22 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
LLM-based vulnerability scanners can provide false assurance when harness logic silently excludes attacker-influenced code, making scan coverage and discarded findings important defensive checks.
Gadi Evron's study of OpenAnt, raptor, and 10 additional LLM-based vulnerability discovery harnesses identifies 13 ways code selection, filtering, and findings handling can suppress real vulnerabilities independently of model capability. Using the harnesses' own code, the study demonstrates a file being excluded before model analysis by six harnesses, five without notice, and also examines content-steered verdicts and prompt injection.