Google disclosed that its PageBreak AI security agent has found more than 500 cross-site scripting vulnerabilities across the company’s first-party web applications. The automated system verifies bugs by running actual exploits against live environments before alerting product teams.
Modern software security is drowning in noise. When automated scanners and large language models hunt for bugs, they routinely flood engineering departments with plausible-sounding hypotheses that fall apart under scrutiny. Google’s Product Security team calls this flood of unverified alerts AI slop
, a problem that often leaves security staff worse off than before.
To combat this verification bottleneck, Google built PageBreak, an autonomous internal AI agent designed to separate genuine security flaws from LLM hallucinations before a human ever sees the report. Information security engineer Michał Bentkowski detailed the system in a September 24 blog post disclosing the project’s milestones.
How PageBreak Tests Exploits in Live Environments
PageBreak operates on a two-stage architecture. Most scans run on Google’s Gemini models, specifically Gemini 3.1 Pro and Gemini 3.5 Flash, though the agent can accommodate other models depending on the scanning requirements. When the AI spots a potential vulnerability, it does not immediately flag it for a product team.
Instead, the agent hands its hypothesis to a purpose-built validator written in conventional software rather than AI. For cross-site scripting, or XSS, the validator injects a JavaScript payload into a running instance of the application, loads the affected page, and checks whether the code actually executes. Depending on the target, XSS allows an attacker to read sensitive data or hijack an active user session.
The validators also check whether injected inputs alter database queries, expose files through path traversal, or trigger remote code execution via sleep delays or outbound DNS requests. Google runs identical seeds across multiple iterations because language models frequently wander down unproductive paths before stumbling on a working exploit in subsequent attempts.
The Gap Between Standard Web Apps and Hardened Frameworks
Across its broader estate of first-party web applications, PageBreak validated more than 500 XSS vulnerabilities, including flaws on sensitive company domains. However, the results changed dramatically when the agent targeted software built on Google’s modern high-assurance web frameworks.
As of September 4, PageBreak identified just two XSS vulnerabilities among hundreds of applications built on those secure frameworks. Google confirmed that both exceptions were confined to internal applications or debug endpoints that lacked full hardening at the time of the scan.
Google is using that stark contrast—more than 500 findings across standard apps versus two in hardened codebases—as empirical proof that safe-by-design architecture can structurally eliminate entire classes of security flaws rather than relying on reactive patching after vulnerabilities emerge.
Infrastructure Advantages and the CodeMender Integration
Independent security researchers cannot easily replicate PageBreak’s scale because the agent relies heavily on internal Google infrastructure. A unified code repository spanning billions of lines allows the AI to trace execution paths across disparate microservices, while live-traffic security data maps page requests directly back to source code.
Unverified hypotheses that fail validation do not disappear entirely. Google retains them internally as training seeds for subsequent scans or as diagnostic feedback when building new validators. To handle the volume of genuine bugs that do pass validation, Google plans to integrate PageBreak more tightly with CodeMender, an automated fix-writing agent previously highlighted by the company.
Читайте также
