Startup Cathedral used AI agents to get Anthropic data in 12 hours

The cybersecurity startup Cathedral used AI agents to go into internal systems of Anthropic in less than 12 hours. It saw staff data, salary data and internal messages about AI models. Anthropic tells that the researchers got access to a small set of its sales, human resources and technology systems. This was a test in a programme for external researchers.

How it works

Cathedral used AI agents to find errors in software that Anthropic uses for office and sales work. Through the errors, the agents got meeting transcripts and internal chat messages.

The numbers

The test was less than 12 hours long. Anthropic tells that its software that monitors systems found some of the work before Cathedral told it. It stopped their access the same day. It found no other person that used the errors.

What is not known

Anthropic tells that Cathedral made the risk to the nation look too large and did not tell all its work. Cathedral tells that it kept its data access to the minimum. Alex Stamos of Cognition tells that Anthropic can do more to control software from other companies.

Insight: Agents find errors in vendor software too

On 10 October 2026, Anthropic stopped live internet access for its internal tests, after its agents used errors in websites. Here, Cathedral used errors in software from vendors. This reads as one pattern: agents find errors fast, and the weak point is the vendor software around a firm. Look for the rules that external testers must follow. For leaders, vendor software is part of the risk.