XBOW Builds Three-Stage Chrome Sandbox Escape With AI Agents in Six Days
Vulnerability Disclosure · 2 sources · Updated 4d ago · First seen Oct 25, 2026
XBOW reported developing a full-chain Chrome exploit in six days by combining three vulnerabilities to turn JavaScript execution in an ordinary browser tab into code execution outside the renderer sandbox. Four researchers directed OpenAI GPT-5.6 Sol agents to search code, test hypotheses and develop exploit stages, while researchers selected targets, challenged findings and integrated the chain. The August 2–7 campaign used roughly $40,000 in compute and earned Google’s first $250,000 Chrome full-chain reward under vulnerability reward rules introduced in April 2026.
The chain combined a V8 Maglev HeapNumber compiler issue, an animated-WebP decoder native-write vulnerability that crossed the V8 sandbox, and a Crashpad memory-safety issue that escaped the renderer sandbox. XBOW demonstrated the chain launching Calculator from a web page in signed Chrome Beta. The research produced two unique CVEs, CVE-2026-85045 and CVE-2026-78952, plus one accepted issue without a verified CVE. Neither CVE has a numerical CNA severity score; Bugflation’s high-severity designation is editorial. The reported development speed underscores the need for browser-security teams to prioritize prompt updates and assess defenses against chained sandbox escapes; the supplied reports do not establish affected versions, patch availability or exploitation in the wild.
