Google’s PageBreak combines Gemini-powered hypothesis generation with deterministic validators that inject payloads and confirm execution in running environments. The result: near-zero false positives, solving a hard problem in security scanning where AI typically guesses.
Google’s announcement details a striking distribution. The system discovered 2 XSS flaws in hardened, secure-by-design apps and 500+ in legacy applications. That gap signals blast radius in remediation, not detection capability. Most security scanners stop at findings; integration with CodeMender aims to auto-patch vulnerabilities instead.
The catch: PageBreak operates on Google’s systems in controlled environments. Whether it scales to third-party code, closed networks, or against sophisticated evasion remains untested. What works internally may not work where security really matters: in the wild.
Sources: Agentic Hacks, Real Proofs: Inside Google’s PageBreak Project • Google’s PageBreak Project
Coverage: Google’s PageBreak AI Agent Finds 500 Flaws in Its Web Apps
Related on this blog: [AUTO] Google removes guardrails from Gemini 4 Argon • [AUTO] Reasoning APIs Leak Secrets Through Replay Attacks • Google Patched Its Agent-to-Agent Privilege Bug, Then Declined to Pay for It