OpenAI blocks Bitcoin security researcher, pushing team to use Chinese AI models instead

On August 9, Rob Hamilton, CEO of Bitcoin custodian AnchorWatch, publicly disclosed that OpenAI's "trust cyber program" had blocked him from continuing AI-powered security analysis on a codebase that had already been responsibly disclosed — even though he had completed the platform's onboarding and KYC processes.

Hamilton leads a volunteer "Bitcoin Red Team" he assembled with developer Calle and roughly 16 other volunteers after a late-July vulnerability disclosure involving Coldcard hardware wallets exposed a flaw that led to the theft of more than 1,000 BTC. The team used frontier AI models to scan as many open-source Bitcoin repositories as possible, as quickly as possible.

The pace was remarkable: in the first 30 hours the team scanned over 390 repositories and recorded 4,962 findings, of which 85 were classified as critical and more than 635 as high-severity, with a reproduction rate of roughly 21%. Total AI compute spending landed between $20,000 and $40,000.

The team used a mix of models including Moonshot AI's Kimi K3, which did most of the heavy lifting, GPT Sol with OpenAI's Cyber Harness, Anthropic's Claude Fable and Opus, and GLM 5.2. Findings were privately disclosed to repository maintainers following standard responsible disclosure practices.

OpenAI was not the only platform to push back — Anthropic also imposed access restrictions early in the project, and both companies revoked access after putting Hamilton through their verification processes. Hamilton described the OpenAI blockage as a policy "local minima" and tagged US officials in his public disclosure to highlight the regulatory challenges compliant researchers face when using American AI tools for legitimate security work.

The practical result was a migration to Kimi K3, which the team found less restrictive. US-based security researchers working to protect Bitcoin infrastructure were pushed toward Chinese AI models because American platforms would not let them do the work.

After the project gained public attention, OpenAI reportedly allowed some access again. The episode underscores the tension between platform abuse-prevention policies and legitimate security research, and how those frictions can reshape which models researchers choose.