AI agents built by Cathedral, a cybersecurity startup valued at $1.4 billion, penetrated Anthropic's internal business systems in less than 12 hours, the startup said in a draft report cited by The Washington Post.

The breach did not reach Claude conversations, source code or model weights. Anthropic confirmed that the researchers accessed what it called "a limited set" of systems covering sales, HR and IT, which store business contact details and basic account records for users, according to Decrypt.

Cathedral was founded by four former Department of Government Efficiency (DOGE) staffers: Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. The company focuses on AI-driven cybersecurity for military uses, according to Gizmodo. It closed a $160 million funding round in July at a $1.4 billion valuation, Decrypt reported.

How did the agents get in?

The agents exploited flaws in outside vendors' administration and sales tools, weaknesses that many companies share, according to Decrypt. Cathedral's draft report said the agents located pathways into:

• transcripts of internal meetings;
• details of salaries and job offers;
• Slack ($WORK) messages concerning Anthropic's advanced models.

Anthropic said the researchers also made several days of failed attempts to breach its software head-on before succeeding through the vendor route, Gizmodo reported.

Cathedral conducted the probe under a program modeled on a bug bounty, an arrangement in which companies pay outside experts for flagging security flaws. Anthropic said Cathedral did not fully follow the program's rules because some of its activity was never disclosed, and it said the startup exaggerated the national security implications, according to Decrypt.

Cathedral's draft report defended its conduct. "Data accesses were kept to the absolute minimum in accordance with security research best practices," the report said.

What was Anthropic's response?

Anthropic said its internal monitoring caught part of the intrusion before Cathedral disclosed it. "Our monitoring flagged some of Cathedral's activity before they reported it to us," the company told The Washington Post.

Anthropic shut off Cathedral's access the same day. The company also pulled the vulnerable tools out of service until the vendors signed off on repairs. The vendors remain unidentified.

The company said it found no evidence that anyone besides Cathedral exploited the vulnerabilities.

Why does the breach worry officials?

A senior Trump administration official, who was not identified, said the report has circulated within government and prompted national security concerns, according to Decrypt.

"If a bunch of kids who have just started a company can do this, then potentially, the Chinese aren't far behind," the official said.

The concern centers on third-party software. Anthropic's most sensitive assets — Claude customer conversations, source code and model weights (the trained parameters that define an AI model's behavior) — stayed out of reach. But the vendor tools that Cathedral compromised are common across corporate America, which extends the exposure well beyond a single AI lab.

The episode also raises questions about how AI agents change offensive security. A probe that once required a team of human penetration testers spanned days of failed direct attempts followed by a sub-12-hour breakthrough through the supply chain, per the timelines both sides described.

What happens next?

Anthropic has not named the vendors or detailed the repairs. Cathedral has not publicly released its final report. The company's draft said the exercise demonstrated how quickly autonomous agents can map and enter corporate infrastructure through third-party software.

For Anthropic, the immediate operational impact appears contained: business contact details and basic account records were exposed, while the core of its AI business stayed sealed. For its enterprise customers, the incident is a reminder that a vendor's security posture is now part of their own attack surface.