Google's AI Security Agent Flags Over 500 XSS Bugs, Only Two in Hardened Apps
Google's internal AI security agent has validated more than 500 cross-site scripting vulnerabilities across the company's first-party web applications, while finding only two such flaws in apps built on its hardened web frameworks — a gap the company is presenting as proof that secure-by-design architecture can withstand relentless automated attack.
The disclosure came September 24 in a blog post by information security engineer Michał Bentkowski of Google's Product Security team. PageBreak began as a pilot in November 2025 and became a full project in January 2026, with a mission to autonomously scale vulnerability discovery while cutting down on manual toil.
Cross-site scripting, or XSS, occurs when an attacker injects a script into a page that another user loads. Depending on the application, the script can read data or take over a victim's logged-in session.
How PageBreak Separates Real Bugs from Hallucinations
What sets PageBreak apart from a conventional large language model scanner is a two-stage architecture. The agent first uses Gemini models — most scans run on Gemini 3.1 Pro and Gemini 3.5 Flash, though other models can be plugged in — to identify suspected flaws. Each candidate is then handed to a purpose-built validator that attempts to fire an actual exploit against a running instance of the application.
For an XSS finding, the validator injects a JavaScript payload, loads the page, and determines whether the script executes. Only after a working exploit succeeds does the issue get flagged as a confirmed bug. Google said this validation step has kept the false-positive rate near zero, and unverified candidates never reach product teams as confirmed vulnerabilities. Instead, they feed back into later scans or help engineers build the validator.
The agent also tests for other attack classes beyond XSS. Its validators check whether injected inputs alter database queries, expose files through path traversal, or trigger code execution. A separate validator examines requests that may reach internal services.
Google runs the same seed over multiple iterations because models can abandon a productive route or pursue an attack path that fails under real application conditions. Repeated attempts give the agent multiple chances to land on a working exploit.
The Hardened Framework Gap
As of September 4, 2026, PageBreak had identified two XSS vulnerabilities among hundreds of applications built on Google's high-assurance web frameworks. Both were confined to internal applications or debug endpoints that were not fully hardened.
The contrast with the broader estate — more than 500 findings across first-party web applications — is the evidence Google is using to argue that safe-by-design frameworks can structurally eliminate entire bug classes rather than requiring patches after the fact.
The agent's reach is amplified by infrastructure that outside researchers cannot replicate. A unified code repository spanning billions of lines lets PageBreak trace execution paths across services. Live-traffic security data maps a page request back to the relevant source code. Existing scanners provide authenticated access to internal sites that would otherwise be difficult to examine.
Google did not name the affected applications or provide a severity breakdown for the 500-plus findings.
Pairing Findings with Fixes
The step is tighter integration with CodeMender, Google's fix-writing agent. The plan is for product teams to review a proposed patch alongside a confirmed vulnerability, reducing the work created by a large volume of verified findings. Google gave no timetable for the integration.
CodeMender was previously cited by Google in May, when its Threat Intelligence Group said it had caught what it believed was the first zero-day exploit created with AI assistance.
The Verification Problem Beyond Google
The challenge of checking AI-generated security reports extends beyond Google's walls. In the crypto software space, the Ethereum Foundation has used separate reviewers to reproduce agent-generated findings, reporting one confirmed flaw in libp2p later disclosed as CVE-2026-34219 while warning that plausible reports can involve unreachable code or attack conditions that do not hold in practice.
An August Bitcoin Red Team scan of 501 open-source projects generated 7,958 findings over 108 hours, but only 24.7% had reproducible proofs at the time. The full tally did not represent 7,958 confirmed exploitable vulnerabilities.
Google has not said whether PageBreak will be made available to external projects.
The disclosure also lands against a backdrop of rising concern AI-enabled cyberattacks. In August, more than 100 organizations including Google, Microsoft, and Anthropic signed an open letter warning that such attacks are becoming more common. Separately, autonomous agents have been found to breach real companies during testing, and an AI agent configured by OpenAI was reported to have accessed an Australian government website without authorization in June.
PageBreak represents the defensive side of that equation: an AI agent tasked with catching the bugs before someone else does.
Once added, BigGo Finance appears first in Google Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
