Google's PageBreak AI Security Agent Finds Over 500 XSS Flaws

By Renato LosioOriginal title: 谷歌公布已发现 500 多个 XSS 漏洞的 AI 安全代理 PageBreak
78AI Score

Google disclosed that PageBreak, its Gemini-powered security agent, has autonomously uncovered over 500 XSS vulnerabilities across its web applications since full deployment in January. The system uses a deterministic verification loop, reporting a flaw only after reproducing it in a live runtime, driving false positives near zero. Google is now integrating PageBreak with agentic security projects like CodeMender to auto-generate fixes.

Full Report

According to a paper published by Google, its AI-driven security agent has found over 500 XSS vulnerabilities across its various web applications since it went fully operational in January of this year. Google PageBreak uses Gemini models to autonomously investigate potential vulnerabilities, but only reports them when a deterministic validator successfully reproduces the vulnerability on a running application, resulting in a near-zero false positive rate.

With PageBreak, each vulnerability hypothesis is tested in a running environment using real exploitation methods, such as executing JavaScript to verify XSS, writing data to a globally readable location to verify path traversal, or triggering an outbound DNS request to verify remote code execution (RCE). Google information security engineer Michał Bentkowski explained:

Early in the development of PageBreak, we made a deliberate decision to prioritize deterministic validation. Rather than merely inferring defects from code patterns, the system closes the loop by validating potential defects in a running environment. This approach brings the false positive rate close to zero, ensuring that product teams are not overburdened by unverified vulnerability reports.

Google also retains non-deterministic detection results internally, providing leads for deeper scans in the future, identifying gaps in its validation capabilities, and guiding the development of new validators. These unverified candidate vulnerabilities are not sent to product teams, allowing them to focus on high-confidence vulnerabilities. Bentkowski wrote:

While deterministic validation is our gold standard, we also recognize that current validators do not yet have the capability to cover all vulnerability types or complex scenarios. This introduces the risk of false negatives—vulnerabilities that might have been detected by non-deterministic validators.

In another article on Google Bug Hunters, "Google's PageBreak Project—Real-World Findings," Bentkowski described three technical cases discovered by PageBreak: an XSS vulnerability, a cache poisoning issue caused by a service misconfiguration, and a path traversal vulnerability. On LinkedIn, Lukas Weichselbaum, a senior information security engineer at Google, emphasized the importance of reducing false positives:

Many AI-generated vulnerability reports sound convincing but ultimately turn out to be wrong, and the teams receiving them have to waste time checking each one. In PageBreak, the model is free to make suggestions, but a finding only counts as valid once a validator confirms it using a real payload. These validators are ordinary code written by security engineers, not generated by the model, and they sit outside the agent loop.

Bentkowski said that PageBreak benefits from Google's engineering environment, particularly Google's monorepo for end-to-end code visibility, security signals that correlate live HTTP traffic with source code, and existing scanning infrastructure that can provide authenticated access to its web applications.

Google's long-term goal is to reduce product team involvement to merely validating AI-generated fixes. To this end, Google is integrating PageBreak with other agentic security projects (such as CodeMender) to automatically generate fixes for the vulnerabilities it discovers.

Currently, these 500+ vulnerabilities have not undergone any independent verification.

Statement: This article is a translation by InfoQ and may not be reproduced without permission.

Original link: https://www.infoq.com/news/2026/10/google-pagebreak-security-agent/