Google's AI Agent Finds 500+ XSS Bugs—and Proves Its Worth by Not Crying Wolf

Google has revealed that its AI-driven security agent, PageBreak, has identified more than 500 cross-site scripting (XSS) vulnerabilities since it began full-scale operations in January. The system combines Gemini AI models with a deterministic validation layer that distinguishes it from many other AI security tools: it only flags bugs that are reproducible against live systems.

How PageBreak Works#

PageBreak operates in two stages. First, the AI models autonomously scan code and propose potential security flaws—examining patterns that might indicate XSS, path traversal, remote code execution, or other attack vectors. Then comes the critical second step: each hypothesis is tested against a running application with a real exploit. For XSS, that means executing JavaScript; for path traversal, writing to a world-readable location; for RCE, triggering an outbound DNS request. Only when the exploit succeeds does the finding move into the formal report queue.

Michał Bentkowski, an information security engineer at Google, describes the philosophy: "Rather than simply hypothesizing bugs based on code patterns, the system closes the loop by verifying potential flaws against running environments. This approach results in a near-zero false positive rate."

The False-Positive Problem in AI Security#

One of the biggest frustrations in AI-assisted security scanning is noise. Unvalidated AI reports can overwhelm product teams with unconfirmed warnings, forcing engineers to spend time chasing phantom bugs. PageBreak sidesteps this by keeping unverified findings internal—using them to retrain validators and spot gaps in coverage—while only escalating high-confidence detections to developers.

This strategy does carry a tradeoff: reliance on deterministic validators may miss complex scenarios or vulnerability types for which validation logic hasn't yet been written. Bentkowski acknowledges the risk of false negatives, yet frames it as preferable to bombarding teams with noise.

Future Direction: Automated Fixes#

Looking ahead, Google is integrating PageBreak with CodeMender, an agentic tool designed to auto-generate patches for discovered vulnerabilities. The ambition is to push human involvement further downstream—moving security teams from validating detections to merely reviewing AI-proposed fixes.

PageBreak also benefits from Google's internal infrastructure: a monorepo offering end-to-end code visibility, security signals linked to live HTTP traffic, and pre-existing scanning systems with authenticated access to production applications. These advantages make the system unusually effective within Google's own environment, though they may be harder to replicate elsewhere.