Security firm XBOW said it built a full-chain Chrome exploit in six days by directing AI agents running OpenAI's publicly available GPT-5.6 Sol model. The chain crossed both of Chrome's nested sandboxes and ran code on Windows when a user simply loaded a web page. Google confirmed and fixed all three bugs, paying XBOW its top $250,000 reward.
The three bugs sat in different parts of Chrome: V8, the JavaScript engine; the animated-WebP image decoder; and Crashpad, the crash handler. The first gave memory control inside the engine, the second broke out of the inner sandbox with a single write, and the third turned that write into full code execution. Two bugs have public identifiers—CVE-2026-85045 for V8 and CVE-2026-78952 for Crashpad—and were fixed in Chrome 152.0.7977.65 and 152.0.7977.82. The WebP bug was fixed in the same cycle, but the article notes a discrepancy: XBOW's announcement labels CVE-2026-85045 as the WebP identifier, while Google and XBOW's own technical write-up assign that CVE to the V8 bug.
XBOW says the AI agents searched Chrome's source code, tested ideas, built proofs of concept, and wrote first drafts of the exploit. Researchers chose targets, supplied context, challenged results, and decided which pieces to join. The awarded chain was built by researcher Brendan Dolan-Gavitt. The six-day run cost about $40,000 in AI charges, according to XBOW's usage logs. The company's first full Chrome chain took two months; Dolan-Gavitt wrote that before LLMs, a team of four would need six to nine months.
The work happened between August 2 and August 7, during Black Hat and DEF CON. XBOW says the bugs were never used against users. The company estimates about 1 billion Windows Chrome users were at risk before the fixes. XBOW founder Oege de Moor and Dolan-Gavitt both framed the speed as a warning for defenders: six days and $40,000 is now enough to produce a Chrome chain.