Ex-DOGE Startup Cathedral Breaches Anthropic in 12 Hours

A months-old cybersecurity startup told Anthropic that its AI agents got inside the company's corporate systems in under 12 hours. They did it by exploiting weaknesses in third-party administrative software, and from there reached meeting transcripts, details of salaries and job offers, and Slack messages discussing Anthropic's advanced AI models, according to the Washington Post report.

The startup is Cathedral, co-founded by Gavin Kliger, until recently the Pentagon's chief data officer. It has raised $160 million at a $1.4 billion valuation and already contracts with three federal agencies. The probe ran under Anthropic's program inviting outside researchers to report vulnerabilities.

In its own description, Cathedral said its "AI agents identified paths into internal systems in less than 12 hours by exploiting weaknesses in third-party administrative software." Anthropic's reply narrowed the scope: Cathedral touched "a limited set" of sales, HR and IT systems containing business contact details and basic user account records, while "customer prompts, Claude responses, source code, and model weights remained inaccessible." The company added that "Our monitoring flagged some of Cathedral's activity before they reported it to us, and we shut off their access the same day," then disabled the affected software until vendors shipped fixes.

A senior Trump administration official pushed a sharper framing to the Post: "If a bunch of kids who have just started a company can do this, then potentially, the Chinese aren't far behind." Anthropic has pushed back on the national-security read of what its monitoring caught and shut down inside a day.

The backdrop is awkward. Kliger was involved while at the Pentagon in the department's legal fight with Anthropic over how Claude could be used by the military, a dispute one write-up notes went 2-1 in the Pentagon's favor at the appeals court on September 25. Cathedral is the third story on agent-driven security we have tracked this week, after a report that most enabled AI agents go unused and lack DLP on sensitive connections and a case where a planted AGENTS.md file hijacked Codex in 90% of tests.