Scanner Dependency Alerts in Code Review: Blind Judges Find No Fully Real Issues

A developer tested whether vulnerability findings from dependency scanners could go directly into an AI code review report without a model verifying them. The reviewer, a Claude Code plugin for reviewing whole backends, runs per-language static tools such as govulncheck, cve-lite and osv-scanner in a Docker image. In the early setup, a model read every advisory and a verifier checked the findings, which cost about 21% of total run cost and sometimes dropped scanner-reported advisories. The replacement used fixed rules that tiered findings by whether the vulnerable code was called or imported, skipping the verifier. Blind Opus judges then reviewed the findings across four private projects, and of 10 distinct issues none was fully real, falling to 0 of 5 after a fix that halved the noise.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.



Discussion (0)
Log in to join the discussion and vote.
Log in