Google Discloses PageBreak AI Security Agent That Found 500+ XSS Bugs

Google has disclosed details of PageBreak, an AI-powered security agent that has uncovered more than 500 cross-site scripting (XSS) vulnerabilities across the company's web applications since it entered full operation in January. The disclosure comes through a paper published by Google and a separate piece on Google Bug Hunters by Michał Bentkowski, an information security engineer at Google, with additional commentary from Lukas Weichselbaum, a senior staff information security engineer there, reported by InfoQ's Renato Losio.
PageBreak runs on Gemini models and works autonomously to investigate suspected vulnerabilities. The defining design choice is what happens before anything reaches a product team: a suspected bug is only treated as a finding once a deterministic validator has successfully reproduced the exploit against a running application. Each hypothesis is exercised against a live environment with a real payload — executing JavaScript for XSS, writing to a world-readable location for path traversal, or triggering outbound DNS for remote code execution. In practice that turns a probabilistic language-model guess into an observable, reproducible event, which is a much stronger signal than a code-review-style suggestion.
Google credits the approach with a near-zero false-positive rate. Bentkowski said the decision to prioritize deterministic validation was deliberate from early in PageBreak's development, with the aim of not burying product teams in unverified reports. Weichselbaum made a related point on LinkedIn, noting that many AI-generated bug reports read convincingly but are wrong, and that teams receiving them lose time checking each one. He described the architecture as one where the model may propose whatever it likes, but a finding only counts once a validator confirms it with a real payload. The validators, he said, are ordinary code written by security engineers rather than by the model, and they sit outside the agent loop. That separation matters: it means the model cannot validate its own output, which limits a common failure mode in agentic systems where the same component both proposes and judges.
Google does keep non-deterministic findings internally. Unverified candidates are not sent to product teams; instead they are used to seed deeper future scans, to identify gaps in validation coverage, and to guide work on new validators. Bentkowski acknowledged a trade-off here: because deterministic validation is treated as the gold standard, validators that cannot yet cover certain vulnerability types or complex scenarios create a risk of false negatives — real bugs that a looser, non-deterministic check might have surfaced. This suggests a system that is optimized for precision over recall, which is a reasonable stance when the output is routed to busy engineering teams, though it does mean the headline vulnerability count is not a ceiling on what is actually present.
Bentkowski attributes part of PageBreak's effectiveness to Google's engineering environment. Specifically, he cites the monorepo's end-to-end code visibility, security signals that connect live HTTP traffic back to source code, and existing scanning infrastructure that provides authenticated access to Google's web applications. A likely reading is that these conditions are doing substantial work: authenticated access and traffic-to-source linkage are precisely the things that let an agent construct and replay real exploits rather than reason abstractly about patterns. Teams without that plumbing would probably need to build equivalent machinery before a similar agent could operate at this precision. The separate Google Bug Hunters article walks through three technical cases found by PageBreak: an XSS vulnerability, a cache-poisoning issue stemming from a service misconfiguration, and a path-traversal vulnerability.
Longer term, Google says it wants to reduce product-team involvement to validating AI-generated fixes, and it is integrating PageBreak with other agentic security efforts such as CodeMender to automatically generate fixes for the vulnerabilities it finds. The reported 500+ figure has not been independently validated.
Why it matters: Security and platform teams evaluating AI-assisted bug hunting now have a concrete architecture to weigh — let a model generate candidates freely, but gate what reaches humans behind reproducible exploits — and the emphasis on precision over recall is a deliberate trade-off others can copy or contest. The dependence on authenticated access, traffic-to-source correlation and a monorepo-style code view suggests, as inference, that reproducing these results outside a large, tightly instrumented organization will be harder than the headline number implies. Developers working on agentic systems more broadly may also read the validator-outside-the-loop pattern as a reusable guardrail for any agent whose output triggers human or automated action.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
According to a paper published by Google, its AI-powered security agent has uncovered more than 500 XSS vulnerabilities across its web applications since entering full operation in January. PageBreak uses Gemini models to autonomously investigate potential vulnerabilities, but reports them only after deterministic validators successfully reproduce the exploit against a running application. By Renato Losio