Google’s internal PageBreak security agent has found and validated more than 500 cross-site scripting vulnerabilities across the company’s first-party web applications, according to details shared by Google and security experts.
The important design choice is validation. PageBreak uses AI to trace potentially unsafe data flows and generate hypotheses, then hands each candidate to a deterministic validator that attempts the exploit against a running application. A finding is promoted only when the validator produces evidence.
That approach aims to avoid the flood of speculative alerts common in static analysis and model-generated reviews. A near-zero false-positive rate, however, applies only to issues the validators can test. It does not show that the system finds every vulnerability or that its test environment perfectly matches production.
PageBreak points toward a practical use of agentic security: automate repeatable actions, observe concrete results and keep evidence attached to the finding. Enterprises evaluating similar tools should measure validated coverage and false negatives, not just the number of bugs reported.
