Google PageBreak Finds More Than 500 XSS Vulnerabilities
Google reports that its Gemini-powered PageBreak security agent has found more than 500 XSS vulnerabilities across its internal web applications since January. PageBreak tests hypotheses in running applications and reports findings only when deterministic validators reproduce them, which Google says keeps false positives near zero. The company is integrating PageBreak with agentic security projects such as CodeMender to automate remediation generation.
PageBreak demonstrates a workflow in which an AI security agent investigates vulnerabilities while conventional validators confirm them before reports reach product teams. However, InfoQ says the 500-plus findings have not been independently validated, and Google acknowledges that limited validator coverage can produce false negatives.

根据谷歌发表的一篇论文,其人工智能驱动的安全代理自今年 1 月全面投入运行以来,已在旗下各类 Web 应用中发现了 500 多个 XSS 漏洞"。Google PageBreak 利用 Gemini 模型自主调查潜在漏洞,但仅在确定性验证器在运行中的应用程序上成功复现该漏洞时才会上报,因此误报率几乎为零。
借助 PageBreak,每个漏洞假设都会在运行环境中通过真实的漏洞利用方式进行测试,例如执行 JavaScript 以验证 XSS、向一个全局可读的位置写入数据以验证路径遍历,或触发外向 DNS 请求以验证远程代码执行(RCE)。谷歌信息安全工程师 Michał Bentkowski" 解释道:
在 PageBreak 的开发初期,我们就刻意决定优先考虑确定性验证。该系统并非仅仅根据代码模式推测缺陷,而是通过在运行环境中验证潜在缺陷来实现闭环。这种方法使误报率接近于零,确保不会因未经验证的漏洞报告而给产品团队造成过重的负担。