RESEARCHERS at XBOW used a public OpenAI model to assemble a Chrome full-chain exploit in six days, chaining three then-unknown Chrome bugs into a single attack that could run code on Windows simply by loading a webpage. The team deployed AI agents running GPT-5.6 Sol to search Chrome’s source, test ideas, build proofs of concept, and draft the exploit, while human researchers provided targets, context, evaluation, and final assembly.
Google later confirmed all three bugs and fixed them; Chrome’s VRP paid XBOW a top prize of $250,000 for a complete Chrome exploit. XBOW notes the six‑day sprint cost about $40,000 in AI charges.
The three bugs sit in different components: the V8 JavaScript engine, the Animated-WebP image decoder, and the Crashpad crash handler. The V8 bug, CVE-2026-85045, and the Crashpad bug, CVE-2026-78952, were fixed in Chrome 152.0.7977.82 (Sep 3) and 152.0.7977.65 (Aug 25) respectively. The WebP-stage bug had no public identifier listed, but Chrome’s cycle addressed it in the 152 series.
Google and XBOW note both identified bugs were high severity and required a user to load a page; the WebP bug’s exact CVE designation is not publicly confirmed. Before fixes shipped, XBOW estimated about 1 billion Windows Chrome users were at risk. Defenders are urged to ensure Chrome is updated to 152.0.7977.82 or later and to relaunch the browser to verify the version.