AI code's security bugs are rarely bad code -- they're missing code

RyzeBlaziken · reddit · 2026-09-17

A developer reviewing AI-generated PRs for a security tool identifies a failure mode: LLMs write what you asked well, but often skip implicit security checks nobody explicitly requested. Real examples: a calendar API endpoint with no ownership check on IDs, and a signing page returning a recipient's PII before verifying authorization. In both cases the existing code was ordinary -- the vulnerability is a branch never written. The author asks whether others see the same omission-vs-commission split and whether research exists on surfacing implicit requirements.

Original post →

More from coding & agent

coding & agent channel →