Adversarial Code Assurance as a Way Back to the Paved Road
summary
Claims that AI coding agents introduce logic flaws like IDOR and broken authorization not caught by classic scanners and that DryRun runs 600,000 code reviews a month and finds 87% agent PRs had security issues in a March test. Walks through a concrete tenant ID (IDOR) bug and shows adversarial multi‑agent review using security context graphs built from Git history, exploit‑chaining analysis, and a paved‑road workflow so agents challenge, repair, and merge safely; one in five PRs changes security posture.
AI-written summary · about
