AI agents cracked Anthropic in under 12 hours

Cathedral, a cybersecurity startup that is only a few months old, says its AI agents found a way into Anthropic's internal systems in under 12 hours. The Washington Post broke the story, citing a draft of Cathedral's own report.

The interesting part is how. The agents did not get through Claude or Anthropic's own code: according to Anthropic, as quoted by Fast Company, several days of trying to break the company's software went nowhere. So they turned to third-party sales and admin tools, the ordinary business software many other companies run too. Through that, Cathedral says, it reached meeting transcripts, salary and job offer details, and Slack messages where staff discussed advanced models.

Anthropic confirmed the access but paints a smaller picture. It says the researchers got into a limited set of sales, HR and IT systems holding contact details for some business customers and partners, plus basic account records. User conversations with Claude, source code and model weights were not touched. Anthropic's monitoring flagged some of the activity before Cathedral reported it, access was cut the same day, and the affected software stayed off until the vendors shipped fixes. The vendors have not been named.

On paper this was a test, not an attack: it ran through a program that works like a bug bounty, where a company pays outside researchers for the holes they find. The two sides are already arguing, though. Anthropic says Cathedral overstated the national security angle and broke the program's rules by not disclosing all of its activity. Cathedral says it kept data access to what responsible research requires.

The team makes it spicier. Cathedral was founded by four former staffers of DOGE, the cost cutting unit Elon Musk ran in the Trump administration. In July it raised 160 million dollars at a 1.4 billion valuation, with Andreessen Horowitz and Sequoia among the backers. A senior administration official told the Post, speaking anonymously, that the report is circulating in government: if "a bunch of kids who have just started a company" can do this, Chinese hackers may not be far behind.

The takeaway is simple. One of the world's top AI labs was opened not through its model but through boring business software, and the ones who did it were AI agents too.