Google has disclosed its AI-powered security agent PageBreak, which identified over 500 cross-site scripting (XSS) vulnerabilities across its web applications since entering full operation in January, according to a paper published by the company. The agent uses Gemini models to autonomously investigate potential flaws, but reports only those that deterministic validators can successfully reproduce against live applications, resulting in a near-zero false positive rate.
Google's paper explains that PageBreak combines AI-generated hypotheses with deterministic validators that reproduce exploits against running applications. Michał Bentkowski, information security engineer at Google, wrote: "Early on in the development of PageBreak, we made a conscious decision to prioritize deterministic validation. 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, ensuring that we avoid overloading product teams with unverified vulnerability reports."
Bentkowski noted that non-deterministic findings are retained internally to seed deeper future scans and guide the development of new validators, but are not sent to product teams. "While deterministic validation is our gold standard, we are aware that our validators are currently lacking in capabilities to cover all vulnerability types or complex scenarios," he added.
Lukas Weichselbaum, senior staff information security engineer at Google, stressed the problem of AI-generated false positives. "A lot of AI-generated bug reports sound convincing but turn out to be wrong, and the teams that get them waste time checking each one. In PageBreak, the model can suggest whatever it likes, but a finding only counts once a validator confirms it with a real payload. The validators are plain code written by security engineers, not by the model and live outside the agent loop."
PageBreak leverages Google's internal engineering environment, including the company's monorepo for end-to-end code visibility, security signals linking live HTTP traffic to source code, and existing scanning infrastructure that provides authenticated access to its web applications.
The agent has already uncovered specific flaws including an XSS vulnerability, a cache-poisoning issue caused by a service misconfiguration, and a path-traversal vulnerability. Google is integrating PageBreak with other agentic security initiatives such as CodeMender to automatically generate fixes for discovered vulnerabilities, with the longer-term goal of reducing product team involvement to validating AI-generated fixes. The company has not provided independent validation of the 500+ vulnerabilities.
Analysis
Why This Matters
- PageBreak shows a practical way to use AI for security without overwhelming teams with false alarms, a persistent challenge in the industry.
- The agent's success could accelerate adoption of autonomous security testing in other organisations, especially those with large codebases.
- The planned integration with automatic fix generation points toward a future where AI not only finds but also patches vulnerabilities.
Background
Automated vulnerability scanning has been a cornerstone of web security for years, but tools often produce high false positive rates that drain engineering resources. PageBreak addresses this by combining AI hypotheses with deterministic verification, ensuring only confirmed bugs are escalated. The system is built on Google's existing infrastructure, including its monorepo and security signals, which may not be easily replicable elsewhere.
What to Watch
- Google's integration of PageBreak with CodeMender for automated patch generation, planned as a next step.
- Whether the system will be extended to cover a wider range of vulnerability types beyond XSS.
- The potential for other large tech firms to develop similar AI agents with deterministic validation.
Checked
12 claims in this story checked against 1 source: 10 supported.
Show each claim
- SupportedPageBreak found more than 500 XSS vulnerabilities.InfoQ: “its AI-powered security agent has uncovered more than 500 XSS vulnerabilities”
- SupportedPageBreak's deterministic validation results in a near-zero false-positive rate.InfoQ: “reports them only after deterministic validators successfully reproduce the exploit against a running application, resulting in a near-zero false-positive rate.”
- SupportedPageBreak found those vulnerabilities across Google's web applications after entering full operation in January.InfoQ: “has uncovered more than 500 XSS vulnerabilities across its web applications since entering full operation in January.”
- SupportedBentkowski said PageBreak verifies potential flaws against running environments rather than merely hypothesizing bugs from code patterns.InfoQ: “Rather than simply hypothesizing bugs based on code patterns, the system closes the loop by verifying potential flaws against running environments.”
- Partly supportedBentkowski said non-deterministic findings are retained internally to seed future scans and guide new validators, and are not sent to product teams.InfoQ: “Google also retains non-deterministic findings internally to seed deeper future scans, identify gaps in its validation capabilities, and guide the development of new validators. These unverified candidates are not sent to product teams”
- SupportedBentkowski said the validators lack the capability to cover all vulnerability types or complex scenarios.InfoQ: “While deterministic validation is our gold standard, we are aware that our validators are currently lacking in capabilities to cover all vulnerability types or complex scenarios.”
- SupportedWeichselbaum said AI-generated bug reports can sound convincing but be wrong, wasting teams' time.InfoQ: “A lot of AI-generated bug reports sound convincing but turn out to be wrong, and the teams that get them waste time checking each one.”
- SupportedPageBreak uses Gemini models to autonomously investigate potential vulnerabilities.InfoQ: “Google PageBreak uses Gemini models to autonomously investigate potential vulnerabilities”
- SupportedPageBreak uncovered an XSS vulnerability, a cache-poisoning issue caused by a service misconfiguration, and a path-traversal vulnerability.InfoQ: “three technical cases found by PageBreak: an XSS vulnerability, a cache-poisoning issue caused by a service misconfiguration, and a path-traversal vulnerability.”
- SupportedPageBreak uses Google's monorepo, security signals connecting live HTTP traffic to source code, and scanning infrastructure providing authenticated access to web applications.InfoQ: “Google’s monorepo end-to-end code visibility, security signals that connect live HTTP traffic to source code, and existing scanning infrastructure that provides authenticated access to its web applications.”
- SupportedGoogle is integrating PageBreak with CodeMender to generate fixes, with the longer-term goal of reducing product-team involvement to validating AI-generated fixes.InfoQ: “With the longer-term goal of reducing product-team involvement to validating AI-generated fixes, Google is integrating PageBreak with other agentic security initiatives such as CodeMender to automatically generate fixes”
- Partly supportedGoogle has not provided independent validation of the more than 500 vulnerabilities.InfoQ: “No independent validation of the 500+ vulnerabilities has been provided.”
