AI agents built a $250K Chrome exploit chain in just six days

Security firm XBOW says its AI agents, running OpenAI's publicly available GPT-5.6 Sol, built a full-chain Chrome exploit in six days (August 2–7, during Black Hat and DEF CON). Compute cost was about $40,000. Google paid the $250,000 top reward under its Chrome Vulnerability Reward Program. XBOW calls it the first payout under Google's 2026 rules.

The chain links three previously unknown bugs in different parts of Chrome: a V8 Maglev compiler flaw, an out-of-bounds write in the animated WebP decoder that escapes the V8 sandbox, and a memory bug in the Crashpad crash handler on Windows. Together they run code outside the browser on a Windows PC when the user only loads a web page. Google has confirmed and fixed all three. XBOW says none was used against real users. Relaunch Chrome so the update takes effect.

This was not fully autonomous. Four human researchers picked the targets, checked the agents' output and made the key call to build the chain from swappable parts. That decision mattered when Google said one bug in the first chain was already known, because the team could replace that part and keep the rest. The agents searched source code, tested ideas and wrote proof-of-concept code and first drafts. As XBOW's Brendan Dolan-Gavitt put it: "Defend accordingly."

Why it matters: exploits for the world's most hardened browser used to take elite teams months; a public model and $40K just did it in under a week.