AI agents got inside Anthropic in 12 hours
Security startup Cathedral used AI agents to get into Anthropic's internal systems in less than 12 hours, The Washington Post reported on the 26th of October 2026. Anthropic confirmed limited access but disputes what the testers saw.
Cathedral's AI agents inside AnthropicSecurity startup Cathedral used AI agents to get into Anthropic's systems in under 12 hours, The Washington Post reported on the 26th of October 2026. Anthropic, the maker of the Claude chatbot, confirmed that Cathedral reached "a limited set" of its sales, HR and IT systems.
Cathedral tested Anthropic under its bug bounty program, a scheme that pays outside experts to find security holes. The two companies disagree on what the testers saw.
How Cathedral got in
The agents used flaws in office software that Anthropic and many other firms use, the Post wrote. The makers were not named.
Cathedral told Anthropic and the software makers about the flaws earlier this month, its draft report says.
What Cathedral says it saw
The draft report says the agents found paths to meeting transcripts, salary and job offer details, and Slack messages. Slack is a work chat app, and the messages discussed Anthropic's advanced AI models.
The report says data access was kept "to the absolute minimum".
What Anthropic says
Anthropic said the testers reached business contact details for some customers and partners, and some basic user account records. It said they could not reach chats with Claude, its code or its model weights, the files that make up the AI.
Anthropic said its monitoring flagged some of the activity, and it shut off access that day. It also said Cathedral did not report all of its activity, as the bounty rules require.
Who Cathedral is
Cathedral was founded by former staff of DOGE, the US government cost cutting team. It has raised 160 million dollars and is valued at 1.4 billion dollars, the Post reported.
Below we break down the claims
Cathedral's 160 million dollars of funding is about 11 percent of its 1.4 billion dollar value, by our arithmetic. Its way in took less than half a day.
What is confirmed and what is not
Anthropic confirmed the access but says Cathedral overstated what it means for national security. Anthropic said it found no sign that anyone else used the flaws.
Key facts
Cathedral's AI agents got in within 12 hours.
Anthropic confirmed access to sales, HR and IT systems.
Anthropic said its code and model weights were not reached.
Anthropic shut off the access the same day.