Google revealed that an internal AI proxy PageBreak developed by its product security team has identified over 500 vulnerabilities in the company's own Web applications. Unlike common automated scanning tools, this system generates reports only after verifying the exploitation of vulnerabilities in a real-running environment, which Google claims results in an false positive rate that is close to zero.

From a pilot project to a formal project

According to a blog post published by Google on September 24, PageBreak will launch a pilot program in November 2025 and transition to a formal project in January 2026. The goal is to expand the scale of vulnerability discovery while reducing manual troubleshooting efforts.

Google stated that its security team has faced an increasingly common problem over the past two years: a large number of vulnerability reports generated by AI seem legitimate, but in reality, they do not exist, which increases the cost of manual verification. The focus of PageBreak's design is to first verify whether the vulnerabilities are truly exploitable before handing them over for manual processing.

Verify first, then report.

This system operates based on the Gemini model. When PageBreak identifies suspicious issues, it will pass the findings to a dedicated verification module. This module will attempt to exploit the vulnerabilities in the online running copy of the application. Only upon successful verification will a formal report be generated.

Google stated that more than 500 XSS vulnerabilities have been discovered in the company's first-party Web applications. Such vulnerabilities could be used to hijack logged-in sessions, steal data, or perform operations on behalf of users.

• Pilot launch time: November 2025
• Official project timeline: January 2026
• Number of vulnerabilities found: over 500

Only two vulnerabilities were found in the new framework.

Google also mentioned that in applications built using its newer "high-security" Web framework, only 2 vulnerabilities were found. Based on this, Google believes that enhancing software security from the development stage is more effective than continuously fixing vulnerabilities afterwards.

The article also mentions that Google possesses conditions that most companies do not have, including a unified code repository and years of internal scanning infrastructure, therefore this approach is not easily replicable by small enterprises.

Next step: integrate automatic repair.

Google plans to connect PageBreak with the automatic patch generation proxy CodeMender. In this way, verified vulnerabilities may come with repair suggestions when submitted to engineers, who will mainly be responsible for reviewing and approving them, rather than having to write patches from scratch.

Additional information:It was mentioned in the text that Google had previously patched a vulnerability in its own AI coding tool, which could have been exploited by attackers to execute malicious code.