AI Hacked Chrome in 6 Days for $40K — and Google Paid $250K for the Receipt
XBOW chained three unknown Chrome bugs into a full takeover using a public OpenAI model — work that used to take elite teams 6–9 months. The offense-defense race just changed forever.
OzeroPublished October 17, 20265 min read
On October 13, AI pentesting startup XBOW published the write-up the security world had been dreading: its agents — running **OpenAI’s publicly available GPT-5.6 Sol**, the same model anyone can call via API — built a full-chain Chrome exploit in **six days, for about $40,000 in compute**. Three then-unknown bugs, chained together, turned a single web page load into code execution on Windows. Google confirmed the chain, fixed all three flaws, and paid **$250,000** — its top bounty, the first full-chain reward under the 2026 rules, and the first ever paid to an autonomous security company.
What a "full chain" actually means
Modern browsers are fortresses with locked inner rooms. A web page runs inside a sandbox, so even a bug should stay trapped. A full chain picks every lock on the way out. XBOW’s chain used three bugs: one in V8 (Chrome’s JavaScript engine), one in the animated WebP image decoder (to write outside V8’s protected memory), and one in Crashpad, the crash-report collector — which runs *outside* the sandbox. In the demo, simply loading a web page launched the Windows Calculator: the classic proof an attacker could run anything. No click, no download.
- Bug 1 — V8 JavaScript engine: the entry point inside the renderer sandbox
- Bug 2 — Animated WebP decoder: a write outside V8’s protected memory area
- Bug 3 — Crashpad crash handler: runs outside the sandbox → code execution with user rights
- Result: remote code execution on Windows from loading one page; ~1B Windows Chrome users theoretically affected
Six days — with important asterisks
The headline number deserves the context XBOW itself provided. The awarded chain was built August 2–7 in Las Vegas during Black Hat and DEF CON — but the project started in January, and XBOW’s *first* chain took two months before being disqualified (Google already knew one of its bugs). The six-day run reused techniques from that earlier work. And this was not a one-shot prompt: agents did the source reading, hypothesis testing, and exploit drafting, while four human researchers — Brendan Dolan-Gavitt, Andres Luksenberg, Baul Lee, Willem Pinckaers — chose targets, supplied context, and decided how to chain the pieces. Still, the compression is staggering: pre-LLM, Dolan-Gavitt estimates a four-person team needed **6–9 months** for comparable work.
Why Google paid the maximum
Google hands out at most four full-chain rewards a year. The $250,000 was approved by its vulnerability rewards panel on August 31, under Chrome VRP rules revised in April 2026. The payment is also a signal: Google wants this research happening in the open, reported privately, and fixed before it leaks — because the same economics now apply to attackers. A $40,000 exploit budget is within reach of ransomware crews, not just nation-states. XBOW’s own advice, quoted across the coverage: defend accordingly.
The part nobody should miss
There is a second story inside this one. The model was not special — GPT-5.6 Sol, public API, no restricted weights. The scaffolding was: agent harnesses, human expertise, and months of groundwork. That means the capability is *diffusing*, not concentrated. Every improvement in general agent tooling is now, automatically, an improvement in offensive capability. The FTC subpoenaed Anthropic and OpenAI on October 15 partly over incidents like this — models reportedly hacking websites in testing. The policy conversation and the technical reality just collided in public.
Ozero’s take
As a crypto user, this is the story that actually keeps me up at night — more than any price chart. Your seed phrase lives behind a browser. Your exchange session lives behind a browser. A full chain like this, in the wrong hands, does not need your password: it needs you to load a page. The bugs are fixed — update Chrome to 152.0.7977.82 or later and *relaunch* (the patch only takes effect on restart) — but the trend line is the real message. Defense now has to move at agent speed: shorter patch windows, hardware wallets for anything that matters, and a healthy paranoia about which links you open. The $250K bounty was the cheapest security purchase Google made all year.
Risks to watch
- This article is educational content based on public reporting — not security or financial advice. Technical details come from XBOW’s published write-up and press coverage; do not attempt to reproduce exploits. Keep your browser updated.
Conclusion
Six days, $40,000, one web page, total compromise — built with a model anyone can rent. XBOW did the responsible thing and told Google first. The uncomfortable question is who does it next, and whether they report it at all. Update your browser, use a hardware wallet for real money, and assume the offense-defense race just got a lot faster.
Primary sources
- Cyberpresso: XBOW used a public OpenAI model to break through every layer of Chrome
- CSO Online: XBOW spent $40K on AI and cracked Google Chrome wide open
- ZeroHunt: XBOW built a Chrome full-chain exploit with a public AI model in six days
- AI Cyber Brief: AI agents built a $250,000 Chrome exploit chain in six days



