XBOW AI Agents Chain CVE-2026-85045, WebP Decoder, and CVE-2026-78952 into $250k Chrome Full-Chain in Six Days

The awarded chain combined a V8 type confusion (CVE-2026-85045), an animated WebP decoder write-what-where primitive fixed in the same 152 cycle, and a Crashpad out-of-bounds write (CVE-2026-78952) that executed outside both Chrome sandboxes. All three bugs were confirmed by Google, patched in 152.0.7977.65 and .82, and had never been observed in the wild. XBOW logged $40,000 in API spend and maintained modular exploit components that allowed rapid substitution when one primitive was found already known to Google.

Pre-LLM teams required four researchers six to nine months for equivalent work; XBOW's first internal chain took two months after a Malta planning meeting in January. The six-day sprint occurred during Black Hat and DEF CON, with agents handling source traversal, PoC generation, and initial exploit drafts while humans selected targets, validated outputs, and assembled the final chain. This pattern matches earlier XBOW internal metrics showing AI compression of reconnaissance and primitive development phases.

The result accelerates the shift from manual to agent-augmented vulnerability discovery. Google now faces compressed disclosure windows and must recalibrate reward structures and patch cadence as public models lower the barrier for full-chain construction. Independent verification of the bugs and reward payment appears in Chrome release notes and XBOW's technical report, while official attribution remains limited to the researcher's name on two CVEs.