AI Agents Built a Working Chrome Attack in Six Days. Google Paid $250,000.
A research team at XBOW used a public OpenAI model to find three unknown Chrome flaws and chain them into a one-click takeover. Google shipped the fix in Chrome 152 alongside hundreds of other security patches.
Key points
- Researchers at XBOW used a publicly available OpenAI model, GPT-5.6 Sol, to direct AI agents that found three unknown Chrome bugs and chained them into a working attack in six days.
- The attack runs code on a Windows machine when the victim loads a web page. No clicking or downloading required.
- Google paid XBOW its top reward, reported as $250,000, and shipped fixes in Chrome 152.
- One bug in the chain, CVE-2026-79282, is a critical use-after-free flaw in ANGLE, the graphics layer Chrome uses to draw web content.
A team at the security firm XBOW has built a full working attack on Google Chrome using AI agents driven by a public OpenAI model, and Google has paid them its top bounty for it.
The researchers chained three previously unknown Chrome bugs into one attack. A victim running Windows only has to open a booby-trapped web page. The attacker's code then runs on their machine.
XBOW says it built the whole thing in six days by pointing AI agents, software that can plan and run tasks on its own, at GPT-5.6 Sol, an OpenAI model anyone can use. The agents hunted for the bugs, wrote the code to trigger them, and welded the pieces into a single chain. We first covered XBOW's AI-driven work back in July; this result is a different order of magnitude.
Google confirmed all three flaws, patched them, and paid XBOW the top reward. The Hacker News reported the figure as $250,000.
What did Google actually fix?
Google shipped the fixes in Chrome 152. One of the bugs in the XBOW chain is listed as CVE-2026-79282, a critical use-after-free in ANGLE. A use-after-free is a memory bug where code keeps referencing memory the program has already discarded, which attackers can exploit to run their own instructions. ANGLE translates web graphics calls for the operating system and has a long history of serious vulnerabilities.
| Detail | Value |
|---|---|
| Fixed version | Chrome 152 |
| Named critical bug in chain | CVE-2026-79282 (use-after-free, ANGLE) |
| Reported bounty to XBOW | $250,000 |
Why does this matter beyond Chrome?
The hard part just got cheaper. Writing a browser exploit chain used to take a skilled human team weeks or months. XBOW's team did it in six days with a model any paying customer can call.
That doesn't mean a bored teenager will produce the next Chrome zero-day tomorrow. XBOW is a professional offensive-security outfit, and its agents run inside tooling the company has spent considerable time building. The judgement worth making is narrower: the floor for serious browser bug-hunting is dropping, and defenders should assume criminal groups are running the same experiments without the courtesy of a bug report at the end. Our earlier story on AI agents and the May 2026 RubyGems attack showed the same playbook used offensively, with no responsible disclosure in sight.
Google's own Threat Analysis Group has spent years warning that commercial spyware vendors buy Chrome chains like this one on the open market. A faster, cheaper production line changes the economics for them too.
Should you worry?
Update Chrome now. Open the browser, click the three-dot menu, choose Help, then About Google Chrome. It'll check for updates and install them. Restart when prompted. You want Chrome 152 or later.
Corporate IT teams running managed Chrome fleets should push 152 immediately. The bugs XBOW reported are patched, but any machine that's skipped the update remains exposed to whatever else was closed in this release.
