Cathedral, a cybersecurity startup founded by four former DOGE staffers, spent several days trying to break into Anthropic’s own systems and failed.
So it pointed its AI agents at the ordinary business software Anthropic buys from other companies. They were inside in under 12 hours.
- They read internal meeting transcripts, salary and job offer details plus Slack messages about Anthropic’s most advanced models.
- Claude conversations, source code and model weights stayed out of reach.
- Anthropic caught part of it, switched the software off the same day and briefed federal officials.
- It says Cathedral broke the rules of its bug bounty program, which pays outsiders to find security holes, by not describing everything it did.
Co-founder Gavin Kliger was the Pentagon’s data chief during its fight with Anthropic over how the government could use Claude.
A senior Trump administration official told the Washington Post that an adversary reaching an AI lab’s internals would be close to getting nuclear weapons grade material. Anthropic says that doesn’t describe what Cathedral actually got.
