Summary
- Google has introduced PageBreak, an AI agent from its Product Security team that has already identified over 500 security flaws.
- PageBreak uniquely verifies vulnerabilities by confirming them with successful exploits in real environments, achieving an almost non-existent false-positive rate.
- The tech giant intends to integrate PageBreak with CodeMender, its automated bug-fixing system.
Google has tasked one of its AI systems with a critical mission: infiltrating its own applications to uncover security vulnerabilities.
On September 24, the company revealed that its Product Security team created an autonomous AI named PageBreak specifically for locating exploitable weaknesses within Google’s web applications, as detailed in a blog post by information security engineer Michał Bentkowski. The concept is straightforward: an AI that reliably identifies real threats without generating false alarms.
“PageBreak is an internal AI agent of Google's Product Security team developed to test the security of our first-party web applications and address this challenge,” Google stated. “Starting as a pilot in November 2025 and transitioning to a fully operational project in January 2026, its goal is to autonomously enhance vulnerability detection while reducing manual effort.”
This initiative addresses a significant issue in security teams worldwide, which have been inundated with "AI slop," a term Google uses to describe the overwhelming number of low-quality, AI-generated bug reports that appear credible but are ultimately inaccurate.
"Distinguishing a genuine, exploitable flaw from a convincing hallucination has become a major challenge," Google noted. When tasked with identifying security vulnerabilities, AI models often generate results, but the authenticity of these findings can be questionable.
PageBreak aims to clarify this uncertainty before any human reviews the findings. Utilizing Google's Gemini models, when PageBreak detects a potential flaw, it forwards the hypothesis to a specialized validator that attempts to exploit the flaw in a live version of the application.
Thus far, PageBreak has discovered over 500 cross-site scripting (XSS) vulnerabilities in Google's web applications, which could enable an attacker to hijack sessions, steal data, or impersonate users on commonly used sites.
When tested against applications built on Google's advanced "high-assurance" web frameworks—designed to eliminate entire classes of bugs—PageBreak identified only two vulnerabilities. This evidence supports Google's assertion that developing secure software from the ground up is more effective than patching vulnerabilities after they are discovered.
The urgency surrounding AI and security has intensified throughout the year. In August, over 100 organizations, including Google, Microsoft, and Anthropic, signed an open letter warning about the rising frequency of AI-driven cyberattacks, following incidents where AI agents from OpenAI and Anthropic breached actual companies during testing. Since then, an AI agent from OpenAI hacked the Australian government, and reports of additional breaches continue to emerge.
In contrast to these breaches, PageBreak represents a proactive approach: rather than facilitating a security compromise, it seeks to identify vulnerabilities before they can be exploited. Google has previously dealt with similar issues, having to patch one of its own AI coding tools after a vulnerability allowed attackers to execute harmful code through it.
Google acknowledges that PageBreak benefits from resources that most companies lack, such as a comprehensive code repository containing billions of lines and extensive internal scanning capabilities, making it difficult for smaller startups to replicate this strategy.
The next phase involves linking PageBreak with CodeMender, Google's automated system for writing patches, enabling confirmed vulnerabilities to be submitted with proposed fixes already included, allowing engineers to review and approve rather than starting from scratch.