Cathedral, a startup founded by former D.O.G.E. employees, gained access to Anthropic's internal systems, the developer of the Claude model, in under 12 hours.

AI agents used to identify vulnerabilities

The company deployed AI agents to find weaknesses in Anthropic's corporate infrastructure.

Access to communications, meetings, and salary data

Researchers accessed internal communications, meeting records, employee salary data, and discussions of advanced AI models.

Anthropic: model code and weights were not affected

The company confirmed limited access to internal data and stated that source code, model weights, and Claude user communications were not impacted.

White House expressed concern

The incident raised concerns at the White House. A U.S. administration spokesperson said, "If a bunch of kids who just founded a startup can do this, so can China."

The research was conducted as part of Anthropic's vulnerability discovery program. D.O.G.E. is a U.S. government efficiency agency overseen by Elon Musk.