Sep 18, 2026 · 19 min
AI agents expose the new fault line in software security
OpenAI Got Hacked By... AI
The episode connects an AI-assisted breach of OpenAI’s private codebase with a broader race to deploy increasingly capable systems in law, biology, and research.
- 1Security researchers used Anthropic’s Claude to access OpenAI’s private GitHub monorepo through a bug bounty program.
- 2OpenAI and Anthropic are expanding AI into legal workflows, parallel-thread projects, and biology research.
- 3Speculation about AI-generated mathematical breakthroughs is helping revive venture interest in ambitious deep-tech sectors.
Don't miss
Security researchers use Anthropic’s Claude to access OpenAI’s private GitHub monorepo through a bug bounty program.
The brief
Brian McCullough frames a week in which AI competition produces an uncomfortable pairing: researchers use Claude to reach OpenAI’s private code, while both labs broaden their commercial and scientific ambitions.
The security story turns on Anthropic’s Claude being used by researchers to access OpenAI’s private GitHub monorepo through a bug bounty program, blurring the line between testing and exposure.
OpenAI’s Astra for Law targets select firms, while Anthropic redesigns Claude Projects around parallel threads and opens a Bay Area biology lab, extending the rivalry beyond chatbots.
The discussion then shifts from present-day products to speculation that AI could generate mathematical breakthroughs, a possibility helping renew venture interest in difficult deep-tech fields.
Taken together, the stories suggest that AI’s next competitive edge may come less from standalone models than from how safely they operate inside consequential institutions.