Google has paid out the first bounty under its newest, richest tier for Chrome vulnerability research — and the researchers who earned it say they built the winning exploit chain largely by directing an off-the-shelf AI model rather than a specialized hacking tool. The security startup XBOW Labs says it used OpenAI's publicly available GPT-5.6 Sol model to construct a full exploit chain against Chrome in six days, collecting the $250,000 reward, according to reporting from Runtime Wire and a separate write-up from ExplainX.
The chain linked flaws in Chrome's V8 JavaScript engine, its animated WebP image decoder, and the Crashpad crash-reporting component to achieve code execution outside the browser's sandbox — the kind of full "browser process" compromise that Google's revamped rewards program, introduced earlier this year, now pays up to $250,000 to see reported responsibly. The program can pay an additional $250,128 for chains that also defeat Chrome's MiraclePtr memory-safety protections, according to coverage of the restructured rules from BleepingComputer, which reported that Google deliberately raised payouts for the hardest-to-find bugs while trimming rewards for flaws that AI tools have made comparatively routine to locate.
Humans chose the targets; the model did the digging
According to XBOW's own account, the work ran from August 2 to August 7 and cost roughly $40,000 in API charges. AI agents searched Chrome's source code, tested hypotheses, and assembled proof-of-concept code, while human researchers selected which components to target, supplied context, and ultimately stitched the pieces into a working chain. XBOW's founder, Oege de Moor, emphasized on social media that the team deliberately avoided a purpose-built offensive model in favor of a general-purpose one.
XBOW didn't use Mythos or a specialized cyber model. Just an off-the-shelf foundation model with a magic on top.
Oege de Moor, XBOW Labs founder
Google reproduced the reported flaws and shipped fixes, according to the coverage. XBOW says it is only the second time Google has issued a full $250,000 Chrome payout, and that just four such awards are expected to go out in 2026 under the current rules.
The episode adds to a growing body of evidence that large language models, even general-purpose ones never trained specifically for offensive security work, can meaningfully speed up vulnerability research when paired with experienced human direction. Separately, a comparable experiment by the research group Hacktron used the same GPT-5.6 Sol Ultra model to reconstruct an exploitable Chrome bug chain directly from public patch commits — a reminder that the techniques cut both ways, since the same agentic workflows that help bug-bounty hunters report flaws responsibly could, in less careful hands, shorten the runway defenders have to patch before an exploit is weaponized.