Startup says its AI agents broke into Anthropic's systems

Cathedral, a new cybersecurity startup, says its AI agents broke into internal systems at Anthropic, the company behind the Claude chatbot. Anthropic confirms the researchers got into a limited set of systems, but says they did not reach Claude users' conversations. The two sides also disagree on how the test was carried out.

What Cathedral says it did

According to a draft report the Washington Post reviewed, Cathedral's AI agents found paths in within 12 hours. Agents are AI programs that carry out tasks on their own. The report says they got in by exploiting weaknesses in third-party business software, meaning tools a company buys from outside vendors. Cathedral says the paths led to meeting transcripts, salaries, job offers and Slack messages.

What Anthropic says

Anthropic confirmed access to a limited set of its sales, human resources and IT systems. That included business contact information for customers and partners, plus some basic user account records. The company says the researchers did not reach Claude users' conversations, source code or AI model weights. Model weights are the core data that makes an AI model work.

Anthropic also says the researchers had earlier spent several days failing to break in directly. The test was part of a program where Anthropic rewards people who find security flaws. The company accuses Cathedral of breaking the program's rules by not disclosing the full extent of its activity. Anthropic did not immediately respond to Gizmodo's request for comment.

Who is behind Cathedral

Cathedral focuses on AI-driven security for military uses. Its founders are former staffers of the now-defunct Department of Government Efficiency: Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. The company has been valued at $1.4 billion and has raised $160 million from investors including Andreessen Horowitz and Sequoia Capital.

Why it matters

An unnamed senior Trump administration official said the episode raised national security concerns in Washington. The fear is that Chinese hackers could abuse similar weaknesses. Officials worry most about spying on AI labs, since stolen trade secrets could help a rival copy their models.

Cathedral's report says weaknesses in ordinary business software can compromise even the most advanced organizations. OpenAI was also recently hit by an ethical hack, a break-in meant to expose flaws, which exposed several employees' ChatGPT accounts.

What to do

• List the outside software tools your business uses for sales, HR and IT.
• Ask each vendor how it handles security updates and who can access your data.
• Limit what each tool can see, so one weak spot opens fewer doors.