Researchers Used a Public OpenAI Model to Build a $250,000 Chrome Full-Chain Exploit in Six Days

Security firm XBOW used OpenAI's publicly available GPT-5.6 Sol model to drive AI agents that assembled a complete Chrome exploit chain in only six days, earning Google's maximum $250,000 bug bounty for a full-chain report. The attack combined three previously unknown flaws — a V8 engine issue tracked as CVE-2026-85045, a bug in Chrome's animated-WebP image decoder, and a Crashpad crash-handler issue tracked as CVE-2026-78952 — to escape both of Chrome's sandbox layers and run arbitrary code on a Windows machine simply by visiting a malicious page, no clicks or downloads needed. Google has already shipped fixes in Chrome 152.0.7977.65 and 152.0.7977.82, and XBOW states the bugs were never exploited against real users, though an estimated billion Windows Chrome users were exposed before the patch. The case highlights how AI-assisted research can compress months of exploit-chain development into days, underscoring the need for users to restart their browsers to apply the latest security update.