XBOW Chains Three Chrome Flaws Using OpenAI Models, Earns $250,000 Bounty

Security company XBOW announced on October 13, 2026 that it developed an exploit chain against Google Chrome by chaining three vulnerabilities, using publicly available OpenAI models. Google paid a 250,000 dollar bounty for the research, and all related vulnerabilities have been fixed.

The attack combined flaws in the JavaScript engine V8, WebP animation image processing, and the crash reporting tool Crashpad. According to XBOW, the chain achieves remote code execution outside Chrome's sandbox on Windows. In a public demonstration, opening a webpage in signed Chrome Beta executed the Windows calculator.

Google's official security advisories identify the V8 race condition as CVE-2026-85045 and the out-of-bounds memory write in Crashpad as CVE-2026-78952, both rated High severity. XBOW researcher Brendan Dolan-Gavitt is credited as the discoverer of the vulnerabilities.

The flaws were fixed in Chrome 152.0.7977.82, released on September 3, and Chrome 152.0.7977.65, released on August 25, according to dailysecu.com.

Six days of AI-assisted research

XBOW said the focused development period for the bounty-winning chain ran from August 2 to August 7, 2026, six days in total. The company used AI agents based on OpenAI's GPT-5.6 Sol model, dividing work across source code exploration, attack hypothesis validation, proof-of-concept and exploit component development. Humans decided what to explore, provided context, verified results and connected the components.

XBOW added that it had been researching Chrome attack chains since early in the year, and that creating the first chain took about two months. API usage fees for the six-day effort came to about 40,000 dollars. Chrome applies both memory isolation inside its JavaScript engine and OS-level isolation of web page processing processes to prevent an attack in one place from spreading.