Google has disclosed that its PageBreak AI security agent has found over 500 cross-site scripting vulnerabilities across the company’s web applications.
Summary
- Google says PageBreak tests suspected flaws against running applications before sending reports to product teams.
- The agent uses Gemini models for most scans and separate tools to confirm whether an exploit works.
- Applications built on Google’s high-assurance frameworks had two XSS findings as of Sep. 4.
- Google plans to connect PageBreak more closely with CodeMender, which generates security fixes.
The Google Product Security team said PageBreak began as a pilot in November 2025 and became a formal project in January 2026. It tests Google’s own web applications and has found cross-site scripting, or XSS, flaws even on sensitive company domains. Google did not identify the affected applications in its announcement.
XSS occurs when an application allows an attacker’s script to run in another user’s browser. Depending on the application and the attacker’s access, the flaw can expose data or let someone act through an affected user’s session. Google reported more than 500 findings across its applications, but did not give a breakdown by product or severity.
How PageBreak confirms suspected flaws
Rather than sending every suspected bug to a product team, PageBreak passes each candidate to a specialized validator. For an XSS finding, the validator inserts a JavaScript payload, loads the affected page, and checks whether the script runs. Google said the validation step has kept the system’s false positive rate close to zero.
The agent can also test other types of flaws. According to Google, its validators check whether an injected input changes a database query, whether an application exposes a file through path traversal, or whether it can be made to execute code. A separate check looks for requests that an application sends to internal services.
Most PageBreak scans use Gemini models, including Gemini 3.1 Pro and Gemini 3.5 Flash, although Google said the agent can work with different models. The validators themselves are not written by the AI agent. Google also runs agents through repeated attempts because a model can follow an unproductive path before finding a workable exploit.
Google built the validation process in response to a problem its security staff had encountered: AI-generated reports can describe convincing attack paths that fail when tested. Under PageBreak’s process, unverified candidates stay within the security team’s workflow. They can guide later scans or help engineers build new validators, but Google said they are not sent to product teams as confirmed bugs.
PageBreak found two flaws in protected applications
Among hundreds of applications built on Google’s high-assurance web frameworks, PageBreak identified two XSS vulnerabilities as of Sep. 4, Google said. Both were confined to internal applications or debug endpoints with gaps in their security protections. The result covers that group of applications; Google’s figure of over 500 findings covers its first-party web applications more generally.
The framework result gives Google a way to test how its application design holds up against repeated scans. PageBreak also has access to company tools that help it inspect applications at scale. Google said its code repository lets the agent follow paths across services, while security data from live web traffic can connect a requested page to the relevant source code. Existing scanners give it authenticated access to internal sites that can be difficult for an outside researcher to examine.
Those resources help explain the scope of Google’s findings without suggesting that another organization could obtain the same results simply by running a Gemini model. PageBreak’s reported count comes from scans of Google’s applications with access to Google’s code, traffic data and testing systems.
Crypto teams face the same verification workload
The problem of checking AI-generated security reports has also surfaced in crypto software. In July, Ethereum Foundation security research described a process in which agents develop potential findings and separate reviewers try to reproduce them. The foundation reported one confirmed flaw in libp2p, later disclosed as CVE-2026-34219, while warning that plausible reports can involve unreachable code or attack conditions that do not hold in practice.
For teams that handle crypto users’ funds, the difference between a candidate issue and a working exploit affects how quickly a report can lead to a fix. An August Bitcoin Red Team scan logged 7,958 findings across 501 open-source projects after 108 hours. At that point, 24.7% of the findings had reproducible proofs; the full tally did not represent 7,958 confirmed exploitable vulnerabilities.
Earlier reporting on crypto bug bounties described a similar review burden. Cosmos Labs co-CEO Barry Plunkett said in April that submissions to its program had risen 900% from the previous year, including both valid and invalid reports. PageBreak is an internal Google tool, and Google has not said it is available to crypto projects.
Google plans to pair findings with fixes
Even after limiting reports to verified findings, Google said its product teams still receive a high volume of security work. PageBreak is therefore working with other Google projects, including CodeMender, an agent that generates bug fixes. Google plans to deepen that connection so product teams can review proposed fixes alongside confirmed vulnerabilities.
Be the first to comment