Google PageBreak AI finds more than 500 web security bugs
Google says its PageBreak agent verifies suspected flaws in live environments before reporting them, aiming to cut false alarms.
By Theo Nakamura · Staff Writer
· 3 min read
Google disclosed its PageBreak AI security-bug system on Sept. 24, saying the internal tool has identified more than 500 cross-site scripting, or XSS, vulnerabilities, a category of web vulnerability, in its first-party web applications. For users and investors watching Google’s AI efforts, the notable detail is the process: Google says PageBreak tests whether a suspected flaw can actually be exploited before reporting it, rather than producing an unverified list of possible problems.
PageBreak is an internal agent built by Google’s Product Security team. It began as a pilot in November 2025 and became a full project in January 2026, according to a post by Google information security engineer Michał Bentkowski. Google says most of its use relies on Gemini models, including Gemini 3.1 Pro and Gemini 3.5 Flash, although the system can work with different models.
How does Google PageBreak AI verify security bugs?
Large language models can flag code that looks risky without establishing that it creates a usable weakness. Google says that noise, including false positives, can add to the workload for product teams reviewing security reports.
PageBreak separates the initial suspicion from verification. When the agent identifies a potential issue, it passes the finding to a specialized validator, software Google says is not AI-written. The validator uses a real payload against a running environment to determine whether the flaw is exploitable.
- For XSS, Google says the validator injects a JavaScript payload and checks whether it executes.
- For SQL injection, it tests whether database queries can be manipulated and checks output or timing.
- For path traversal, it creates a readable file and tests whether the application can retrieve it.
- For remote code execution and server-side request forgery, it looks for evidence such as file creation, delays, or network requests.
Google says this validation approach has produced a near-zero false-positive rate. That is the company’s own performance claim and is not independently verified in the available material. PageBreak keeps unverified candidates out of reports to product teams, Google says, while retaining them to guide later scans and improve its validation tools.
What did PageBreak find in Google’s web apps?
Google reported more than 500 XSS findings across its first-party web applications, including sensitive domains. The company also tested applications built on its high-assurance web frameworks, which are intended to prevent certain web vulnerabilities by default.
As of Sept. 4, Google said PageBreak had found two XSS vulnerabilities across hundreds of applications using those frameworks. Google said both were confined to internal applications or debug endpoints with hardening gaps. The comparison is Google-reported evidence from its own environment, rather than an independently controlled study.
Google says its setup is difficult to reproduce elsewhere. PageBreak draws on a single repository containing billions of lines of code, a system that maps live HTTP traffic to source code, and scanning infrastructure capable of authenticating to nearly every Google web application.
Google said PageBreak is working with other agent-based efforts, including CodeMender, as it handles the volume of confirmed findings. The available material does not establish that either tool is available to outside organizations. Readers can review Google’s technical description of PageBreak for the company’s full account of the system.
This story draws on original reporting from Decrypt.
