Full Breakdown
Cathedral Startup Claims Breach of Anthropic Amid Pentagon Dispute
By Drooid · · How we work
Cathedral’s Access to Anthropic’s Internal Systems
Cathedral, an AI-driven cybersecurity firm founded by former Department of Government Efficiency (DOGE) staffers Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein, reported that its AI agents identified pathways into Anthropic’s internal sales, human-resources and IT systems in less than 12 hours. The draft findings, reviewed by the Washington Post, say the agents exploited weaknesses in third-party administrative software to view employee compensation data, job-offer details, meeting transcripts and Slack messages. Cathedral emphasized that the access was limited to what was necessary for its research and did not include Claude user conversations, source code or model weights.
Legal and Government Context
The incident occurs against a backdrop of a Pentagon-Anthropic legal dispute over the company’s refusal to lift restrictions on certain military uses of its Claude model. On September 25, a federal appeals court ruled 2-1 that the Pentagon could exclude Anthropic from its supply chain. A senior Trump-administration official told the Washington Post that three federal agencies have contracted with Cathedral and are collaborating on potential adversarial-AI scenarios, raising national-security concerns about software vulnerabilities that could be exploited by foreign actors.
Data and Scope of the Breach
Anthropic confirmed that Cathedral’s researchers accessed a “limited collection” of sales, HR and IT records, including basic customer and partner contact information. The company said its monitoring detected the activity, terminated the researchers’ access the same day, and disabled the affected software until vendors addressed the vulnerabilities. Anthropic also reported that by October 2 it had disclosed 6,157 vulnerabilities across 591 open-source projects through its own AI-assisted research.
Official Responses
Anthropic disputed Cathedral’s characterization of the breach as a national-security threat, asserting that the researchers did not fully follow the bug-bounty program’s rules. The firm noted that no evidence showed exploitation of the accessed data by any third party. Cathedral’s report, however, warned that weaknesses in ordinary enterprise software could compromise even the most advanced organizations, framing the incident as a broader wake-up call for U.S. companies.
Implications for AI Security
The episode highlights the tension between incentivizing vulnerability research and enforcing strict testing protocols, especially for AI labs with strategic importance. Government officials’ concern that similar exploits could be leveraged by Chinese hackers underscores the emerging intersection of AI development, cybersecurity, and national-defense policy.