XBOW uses GPT-5.6 Sol to build $250,000 Chrome exploit in six days
Security firm XBOW earned $250,000 from Google’s Chrome bug bounty after its AI‑driven team built a full‑chain exploit in six days, from 2 to 7 August. The researchers, led by Brendan Dolan‑Gavitt, logged about $40,000 in compute costs, which they attribute to running OpenAI’s publicly available GPT‑5.6 Sol model.
Key points
- XBOW earned $250,000 from Google’s Chrome bug bounty for a full‑chain exploit built in six days.
- The exploit chained three Chrome bugs and cost about $40,000 in compute, according to XBOW logs.
- OpenAI’s publicly available GPT‑5.6 Sol was used; the model was later paused over cyber‑risk concerns.
The chain linked three previously unknown Chrome bugs: a Maglev compiler flaw in V8, a WebP decoder issue, and a Crashpad crash‑handler vulnerability (CVE‑2026‑78952). Together they allowed code execution from a simple web page on Windows machines, exposing roughly one billion users, according to XBOW. Google patched the bugs in Chrome 152 updates on 25 August and 3 September. The model GPT‑5.6 Sol was released to everyone in July and briefly paused in August over cyber‑risk concerns. XBOW’s founder Oege de Moor said the model was chosen because it was publicly accessible, and OpenAI’s head of cyber research Clint Gibler said the company wants to help defenders find and fix critical flaws faster.
Dolan‑Gavitt warned that “six days and $40k is not a lot of time or money when the payoff is a Chrome chain,” urging defenders to prepare accordingly.
Model page: GPT-5.6 Sol →
The story so far
2 episodes →- XBOW uses GPT-5.6 Sol to build $250,000 Chrome exploit in six daysthis story
XBOW used GPT-5.6 Sol to build a $250,000 Chrome exploit in six days
thenextweb.com · 17 October 2026
Loading the full article…
This text was published by thenextweb.com and written by Ana Maria Constantin. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Meta releases Muse Wonder model for agents, available to developers · 2 src
- Anthropic consults law firms to develop negotiation norms for AI agents · 4 src
- DeepSeek launches Search and Fetch APIs at $0.75 per 1,000 requests · 1 src
- WhatsApp MCP lets AI agents read personal chats on Cloudflare Workers · 1 src
- Opinion: GPT-6.1 Sol may be the best default model for Codex users · 1 src
Comments
via GitHub Discussions