Google has pulled back the curtain on an internal AI security agent that has quietly identified and validated more than 500 cross-site scripting (XSS) vulnerabilities across the company's own web applications — and its real trick isn't finding bugs at all. It's proving every one of them is genuinely exploitable before a single engineer is bothered about it.
The system is called PageBreak, built by Google's Product Security team. Details were reported this week by Dark Reading and covered in The Hacker News' weekly security recap. PageBreak began as a pilot in November 2025 and expanded into a full project in January 2026, running continuously against Google's first-party web applications, including sensitive domains.
Guessing bugs is easy — proving them is the hard part
AI-driven security scanning has a well-known credibility problem: large language models generate large volumes of plausible-looking vulnerability reports, many of which are hallucinations. Product teams drown in unverified alerts, and real flaws hide in the noise. Google says PageBreak was built specifically to dodge that trap: rather than simply hypothesizing bugs based on code patterns, the system closes the loop by verifying potential flaws against running environments.
The workflow is deliberately split in two. Gemini models analyze applications and generate hypotheses about possible vulnerabilities. But those hypotheses are not treated as confirmed findings. Instead, each candidate is handed to a dedicated deterministic validator that attempts to reproduce the exploit against a running test environment. For cross-site scripting, PageBreak injects a tailored JavaScript payload into the relevant application path, loads the target URL through a browser-like rendering harness, and watches whether the code actually executes in the browser context. Only flaws backed by a successful proof are elevated as confirmed findings — an approach Google says has produced a near-zero false-positive rate.
500 verified flaws — and it's just getting started
So far, PageBreak has identified and validated more than 500 cross-site scripting vulnerabilities across Google's application estate. And the same proof-driven model extends beyond XSS: the validators can check whether SQL queries can be manipulated, whether path traversal can create and read a test file, whether an application can be tricked into making requests to internal services, and whether remote code execution is achievable through timing delays, file writes, or outbound callbacks.
Discovery is racing ahead of repair
The timing is hard to ignore. Google's own Threat Intelligence Group tracked vulnerability disclosures from January 2025 through August 2026 and found that monthly disclosures have doubled, with high-risk ones up 167 percent. AI-discovered flaws are also more dangerous on average: they are roughly twice as likely to enable remote code execution as those found by humans. And the broader industry is straining — Microsoft's July 2026 Patch Tuesday shipped a record 570 patches, including three zero-days.
In other words, AI is finding bugs far faster than anyone is fixing them. That is exactly why Google's bet on exploit-backed, low-noise automation matters: instead of a long list of guesses, engineering teams get a short list of confirmed problems, each one already proven with a working exploit. From guessing bugs to proving them — the arms race between AI attackers and AI defenders just gained a proof-of-concept machine.
By Hamza Faisal
