With PageBreak, each vulnerability hypothesis is tested in a production environment with a real exploit, such as executing JavaScript for XSS, writing to a public location for path traversal, or running outbound DNS for RCE. Michal Bentkowski, an information security engineer at Google, explains:

Early on in PageBreak's development, we made a conscious decision to prioritize deterministic validation. Instead of simply hypothesizing errors based on code patterns, the system closes the loop by testing for potential flaws in running environments. This approach results in a near-zero false positive rate, which allows us to avoid overloading development teams with unverified vulnerability reports.

Google also stores non-deterministic results internally to conduct deeper future scans, identify gaps in its verification capabilities, and guide the development of new validators. These untested candidates are not sent to product teams, focusing their attention on high-confidence vulnerabilities. Bentkovsky writes:

While deterministic validation is our gold standard, we recognize that our validators currently lack the ability to cover all types of vulnerabilities or complex scenarios. This carries the risk of producing false negatives that might otherwise be detected by non-deterministic validators.

In a separate article, "Google PageBreak Project - Real Life Results" on Google Bug Hunters, Bentkowski describes three technical cases discovered by PageBreak: an XSS vulnerability, a cache poisoning issue caused by service misconfiguration, and a path traversal vulnerability. On LinkedIn, Lucas Weichselbaum, senior information security engineer at Google, emphasizes the importance of reducing false positives:

Many AI-generated bug reports sound convincing but turn out to be wrong, and the teams that receive them spend time checking each one. In PageBreak, the model can propose whatever it wants, but the result is only counted if the validator confirms it with a real payload. Validators are simple code written by security engineers, not the model, and live outside the agent loop.

Bentkowski says PageBreak benefits from Google's engineering environment, particularly Google monorepo end-to-end code visibility, security signals that connect live HTTP traffic to source code, and existing crawling infrastructure that provides authenticated access to web applications.

In the long term, Google will integrate PageBreak with other agent-based security initiatives such as CodeMender to automatically generate patches for detected vulnerabilities.

More than 500 vulnerabilities have not been independently verified.