Google's PageBreak AI agent found over 500 XSS flaws in Google's own web apps
Google disclosed that its PageBreak AI agent found over 500 cross-site scripting flaws in its own web apps, confirming each one with a working exploit before reporting it.
Why it matters
The findings concern Google's own properties, and the company named no affected applications and published no severity data, so the public size of the exposure is unclear. PageBreak and CodeMender are internal Google projects that are not offered to outside teams, which leaves the disclosure without a direct read-through to crypto assets or exchanges.
Key facts
Google's Product Security team built PageBreak, an internal AI agent that tests the security of Google's first-party web applications.
Reported by Decrypt
The PageBreak blog post was written by Google information security engineer Michał Bentkowski.
Reported by Decrypt
PageBreak began as a pilot in November 2025 and became a full project in January 2026.
Reported by Decrypt
PageBreak has found more than 500 vulnerabilities across Google's web applications.
500Reported by Decrypt
Crypto.News described the 500-plus findings as cross-site scripting flaws, while Decrypt gave no vulnerability type.
500Disputed figureReported by Decrypt
PageBreak sends a candidate flaw to a specialized validator that inserts a JavaScript payload and loads the affected page, and only reports confirmed exploitable issues to product teams.
Reported by Decrypt
PageBreak uses Gemini models for most scans.
Reported by Crypto.News
What happened
Google's Product Security team disclosed PageBreak, an internal AI agent built to find exploitable vulnerabilities in Google's first-party web applications, in a blog post by information security engineer Michał Bentkowski, Decrypt reported. The agent started as a pilot in November 2025 and moved to a full project in January 2026, according to both outlets. Crypto.News reported that PageBreak has found over 500 cross-site scripting flaws, including on sensitive company domains, while Decrypt put the total at more than 500 bugs without naming a type, and Crypto.News says Google gave no breakdown by product or severity and did not name the affected applications. PageBreak hands each candidate flaw to a validator before a report reaches a product team: Crypto.News says the validator inserts a JavaScript payload into the affected page, and Decrypt, relaying Google's own framing of the design, said the exploit check leaves the agent with a near-zero false-positive rate amid what Google describes as a flood of low-quality AI-generated bug reports. Crypto.News reported that applications built on Google's high-assurance frameworks had two XSS findings as of Sep. 4, a detail only that outlet carries. Google plans to link PageBreak to CodeMender, its automated bug-fixing agent, and neither outlet says that link is in place yet.
How the story developed
- Update
High-assurance framework applications stand at two XSS findings
Crypto.News reports the count as of Sep. 4.
- Official statement
Google discloses PageBreak in a blog post
The company publishes the disclosure written by information security engineer Michał Bentkowski.
- First report
First report by Decrypt
Google Built an AI That Hunts Its Own Security Bugs
- Update
PageBreak starts as a pilot inside Google's Product Security team
The autonomous vulnerability-discovery agent begins as a pilot project.
- Update
PageBreak becomes a full project
The pilot graduates into a fully-fledged project with the goal of scaling vulnerability discovery while cutting manual work.
- Update
Decrypt publishes the PageBreak disclosure
First report of the story on the desk's evidence list.
- Independent corroboration
Independently corroborated by Crypto.News
Google’s PageBreak finds over 500 XSS flaws in its web apps
- Update
Crypto.News reports the XSS tally and confirmation pipeline
Second outlet to cover the story, adding the finding count, the validator mechanism, the Gemini scans and the CodeMender plan.
How coverage built up
Independent sources over time, counted the way the consensus panel counts them: a republication of a wire story does not move the line.
- 25 Sept 2026, 19:16 UTC: Decrypt — 1 independent source, 1 reports
- 25 Sept 2026, 19:38 UTC: Crypto.News — 2 independent sources, 2 reports
Affected entities
Original sources
| Publisher | Report | Role | Published |
|---|---|---|---|
| Decrypt Crypto media | Google Built an AI That Hunts Its Own Security Bugs | Original | |
| Crypto.News Crypto media | Google’s PageBreak finds over 500 XSS flaws in its web apps | Independent |