Cathedral says AI agents breached Anthropic in under 12 hours

AI startup Cathedral has announced that it used autonomous AI agents to break into Anthropic's systems. According to the information available, the agents entered the target company's network through weaknesses in third-party business software. Anthropic confirmed the incident and said access was gained to a limited number of systems.

The episode shows that even advanced AI labs can be exposed through ordinary enterprise software. That one AI company was targeted with autonomous tools built by another AI startup makes the case all the more striking.

How the intrusion unfolded

Anthropic runs two separate bug bounty programs through HackerOne. One targets technical vulnerabilities in the company's systems. The other, which is invite-only, rewards researchers who find universal jailbreaks that get past Claude's safety safeguards. Cathedral is reported to have shared its findings with Anthropic through the bounty program.

According to accounts of the incident, Cathedral's human researchers spent several days trying to get into the system without success. The team then switched to its AI agents, which got in within 12 hours by exploiting flaws in third-party business software.

What data was affected

The agents that gained access to the network reportedly reached meeting transcripts, salary and job offer information, and Slack messages. Employees' conversations with Claude, however, appear not to have been exposed.

Anthropic confirmed that Cathedral accessed a limited portion of its sales, HR and IT systems, and that some contact details for customers and partners, along with basic user account records, were compromised. It stressed that neither source code nor AI model weights were accessed. The company said its monitoring systems detected part of Cathedral's activity before it was notified, and that it shut off the access the same day.

Rule-breaking claim and wider implications

Anthropic is reported to have said that Cathedral violated the program's rules by not disclosing the full scope of its activity. It is unclear how this will affect whether a bounty is paid. The timing is also awkward: Anthropic is reportedly preparing for an IPO, and having sensitive data reached by an outside party, even a white-hat hacker, is not something the company would welcome.

Cathedral is a startup focused on offensive and defensive cyber operations for the US military and government. It was founded by former members of the now-disbanded US DOGE (Department of Government Efficiency), including Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. The company's report says weaknesses in ordinary enterprise software can endanger even the most sophisticated organizations' infrastructure, and that this should serve as a warning to every US company.

Two cybersecurity experts also commented on the incident. One said attackers got in the way they usually do, through suppliers, and that an organization's security maturity is only as strong as the weakest vendor it integrates. The other praised Anthropic's detection and response effort but said the real test is independent verification that the attack paths have been closed.