Anthropic spends its days building the AI agents others worry about. This week, it found out what it feels like to be on the receiving end. Cathedral, a cybersecurity startup founded by former members of the now-defunct Department of Government Efficiency, says its AI agents got into Anthropic's internal systems in under 12 hours — and Anthropic confirms the researchers got in, while disputing what it means.
Per a draft Cathedral report reviewed by the Washington Post and reported by Gizmodo, the agents slipped in through vulnerabilities in third-party business software Anthropic uses for sales and administrative work — not Anthropic's own code. From there, Cathedral says, they reached meeting transcripts, salary and job-offer details, and Slack messages about the company's most advanced models. The Post withheld the vendors' names for security reasons.
Anthropic's version is narrower. The company confirmed the researchers reached "a limited set" of its sales, human resources, and IT systems, including business contact information and basic user account records. But it said they never touched Claude users' conversations, source code, or model weights — the crown jewels. Its monitoring flagged some activity before Cathedral reported it, and access was shut off the same day. The affected software was disabled pending vendor fixes, and Anthropic briefed federal officials.
They also disagree on the rules: Anthropic accused Cathedral of violating its bug-bounty program by not disclosing the full extent of its activity, and says the startup spent days failing to breach its software directly first — and overstated the national-security implications.
Washington is reading it the dramatic way. A senior Trump administration official told the Post: "If a bunch of kids who have just started a company can do this, then potentially, the Chinese aren't far behind." The framing lands because of who Cathedral is: co-founder Gavin Kliger was the Pentagon's chief data officer during its fallout with Anthropic, which banned the company's models from federal systems after it refused to allow their use for autonomous weapons or mass surveillance. Cathedral has raised $160 million at a $1.4 billion valuation, per Dealroom, backed by Andreessen Horowitz and Sequoia, and already holds three federal agency contracts.
Strip away the politics and the scope dispute, and the agreed facts are the story: a frontier AI lab's hardest shell wasn't the problem. The way in was ordinary enterprise software — and AI agents just made finding that door fast.
— via Gizmodo