# Google Says PageBreak Found More Than 500 XSS Bugs

> Google’s internal security agent tests suspected flaws with real exploits before reporting them. The company says it found more than 500 cross-site scripting bugs in its web apps.

Oossa · 2026-10-31 · https://oossa.com/en/google-says-pagebreak-found-more-than-500-xss-bugs

Google says its internal AI security agent, PageBreak, has uncovered more than 500 cross-site scripting vulnerabilities in the company’s web applications since it moved into full operation in January 2026. Cross-site scripting, or XSS, lets an attacker make a website run unwanted code in a visitor’s browser.

The agent uses Gemini models to investigate possible flaws, but a suspected bug is reported to a product team only after a separate, non-AI validator reproduces it against a running application. Google says this approach keeps false positives near zero. InfoQ notes that the findings have not been independently validated.

PageBreak is built for Google’s own first-party web applications, not presented as a tool customers can use. Google says most of its use relies on Gemini 3.1 Pro or Gemini 3.5 Flash, though the system can work with other models.

## How does it check a suspected bug?

Instead of relying only on code patterns, PageBreak passes a possible vulnerability to specialized validators written by security engineers. For an XSS report, a validator injects JavaScript and checks whether it actually runs in a browser. Other checks can try to read a file created in a world-readable location, manipulate a database query, or trigger an outbound network request to test for remote code execution.

Google keeps unverified possibilities out of product teams’ inboxes. It retains them internally to guide later scans and help engineers build validators for vulnerabilities the current checks cannot confirm. That choice has a trade-off: Google acknowledges its validators do not cover every flaw or complicated scenario, so some real bugs may be missed.

## What does Google say it found?

InfoQ reports three examples described by Google security engineer Michał Bentkowski: an XSS bug, cache poisoning linked to a service misconfiguration, and a path-traversal flaw. Google says PageBreak can trace code across its large shared code repository and use security signals from live web traffic to connect website paths to source code. Its existing scanners also provide access to many authenticated web applications.

Google says a separate set of applications built on its high-assurance web frameworks had only two XSS findings across hundreds of applications as of September 4, 2026. Those were in internal applications or debug endpoints with gaps in their hardening. Google presents this as evidence for secure-by-design frameworks; it is not a direct comparison of PageBreak with another scanner.

The company plans to connect PageBreak more closely with automated repair projects, including CodeMender. Its stated goal is for product teams to spend less time investigating reports and more time reviewing proposed fixes. Google has not said when that integration will be complete.

## The facts

- Google says PageBreak found more than 500 XSS vulnerabilities in its first-party web apps.
- Google says the project moved from pilot to full operation in January 2026.
- Most PageBreak usage relies on Gemini 3.1 Pro or Gemini 3.5 Flash, according to Google.
- As of September 4, 2026, Google says it found two XSS bugs across hundreds of apps built on its high-assurance frameworks.

## Why it matters

For Google product teams, requiring a reproducible exploit before a report is passed along could reduce time spent checking false alarms. The company says PageBreak has found more than 500 XSS bugs, but independent validation of that count has not been provided, and Google acknowledges its checks may miss some vulnerabilities.

## Sources & references

1. [Google Discloses PageBreak AI Security Agent That Found 500+ XSS Bugs](https://www.infoq.com/news/2026/10/google-pagebreak-security-agent/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering) – InfoQ, 2026-10-31
2. [Google](https://blog.google/security/agentic-hacks-real-proofs-inside-googles-pagebreak-project/) – Google

Last updated: 2026-10-31