AI-assisted Chrome attack earns XBOW US$250,000 Google bounty
Google has fixed the three vulnerabilities. XBOW says they were not used against Chrome users, but estimates about one billion Windows users were exposed before patching.
XBOW announced on October 13 that it had received a US$250,000 Chrome security bounty from Google after combining three previously unknown flaws into an attack developed with AI assistance. Agents running OpenAI’s GPT-5.6 Sol helped create the exploit, which allowed code to run outside the browser’s sandbox on Windows when a webpage opened.
Google has patched all three vulnerabilities. According to XBOW, none was used against Chrome users. The company described the payment as the first award for a complete Chrome exploit chain under Google’s April 2026 rules.
The successful chain was built by researcher Brendan Dolan-Gavitt and the agents from August 2 to 7. XBOW estimated its computing bill at about US$40,000 from usage logs. That work followed months of research, including an unsuccessful submission: Google already knew about one flaw, so researchers replaced that stage with a backup.
The division of work included agents examining browser code, testing approaches and drafting exploit stages. Researchers supplied context, selected targets, reviewed results and determined how the pieces should connect. XBOW founder and CEO Oege de Moor said the team intentionally used a publicly available model.
The attack crossed two sandbox boundaries through vulnerabilities in the V8 Maglev compiler, animated WebP decoding and Crashpad, Chrome’s crash handler. Google credited Dolan-Gavitt for the Crashpad vulnerability, CVE-2026-78952, and assigned it a high severity rating.
