Google has revealed PageBreak, an internal AI security agent that searches its first-party web applications for exploitable vulnerabilities.
Instead of merely flagging suspicious code patterns, PageBreak tests each hypothesis against a running environment. Specialized validators execute controlled payloads for vulnerability classes including cross-site scripting, SQL injection, path traversal, remote code execution, and server-side request forgery.
Google says this verification loop gives PageBreak a near-zero false-positive rate. At scale, the agent has uncovered more than 500 XSS vulnerabilities across Google web applications, including sensitive domains.
PageBreak is model-flexible, although Google says most runs use Gemini 3.1 Pro or Gemini 3.5 Flash. Unverified findings are retained to improve future scans and validators but are not forwarded to product teams. Google is also connecting PageBreak with CodeMender, its security agent for generating and testing patches.
PageBreak remains an internal Google system rather than a newly available public product.
Why it matters
For builders, PageBreak demonstrates that dependable agents need more than capable models. The important architectural pattern is a closed verification loop: let the model investigate broadly, then require deterministic evidence before its output reaches people or production workflows.
That approach can reduce alert fatigue while making autonomous agents useful for high-stakes engineering work. It also strengthens the defensive response to the growing use of AI in vulnerability discovery: findings become actionable only after the system proves that a suspected flaw can be exploited.