Google's PageBreak, an AI security agent, has identified more than 500 cross-site scripting vulnerabilities in the company's web applications. The tool found only two flaws in hundreds of apps built on Google's high-assurance web frameworks, according to an announcement by the company's Product Security team in September 2026.
PageBreak differs from conventional vulnerability scanners by requiring working exploits against live application instances before flagging a bug. This approach aims to reduce false positives common in AI-assisted security scanning.
How PageBreak Works
The agent primarily uses Gemini 3.1 Pro and Gemini 3.5 Flash models to conduct initial scans. Suspected vulnerabilities are then passed to a purpose-built validator that executes actual payloads against running instances of the target applications. For cross-site scripting attacks, the validator injects JavaScript code, loads the page, and confirms whether the script executes.
Beyond XSS, PageBreak's validators test for database query injection, path traversal leaks, and code execution flaws. Google stated that the tool's false-positive rate is near zero.
Secure-by-Design Results
Both of the two vulnerabilities found in hardened-stack applications were located in internal apps or debug endpoints that were not fully hardened, as of September 4, 2026. Google cited the stark difference between the 500+ findings across its broader web application portfolio and the two in secure-by-design frameworks as evidence that architectural hardening can withstand automated security testing.
Development Timeline
PageBreak began as a pilot program in November 2025 and became a full project in January 2026. The tool was announced by information security engineer Michał Bentkowski.
Integration Plans
Google plans to integrate PageBreak more closely with CodeMender, an agent designed to generate security patches. This integration would allow teams to review proposed fixes alongside confirmed vulnerabilities.
