Startup Cathedral Discovers Vulnerabilities in Anthropic Systems in 12 Hours

In October, the startup Cathedral notified Anthropic of a breach into the internal systems of Claude's developer. This was reported by The Washington Post, citing a draft report. Vulnerabilities were found in third-party software for sales and administrative tasks: AI agents took less than 12 hours to find paths to meeting transcripts, salary data, and Slack messages discussing Anthropic models. The testing was conducted as part of a bug bounty program.

Anthropic confirmed access to parts of the commercial department, HR, and IT systems, closed the loopholes, and notified federal officials. Client correspondence with models, source code, and neural network weights remained inaccessible. According to the company, internal monitoring detected some actions before Cathedral's notification; the researchers were blocked the same day, no other breaches were found, and services were restored after supplier fixes. Anthropic stated that Cathedral exaggerated the significance of the findings for national security and did not fully comply with the program's rules. Cathedral's report states that data access was minimized, and the results show a gap between offensive capabilities and asset protection. Cognition's head of security, Alex Stamos, called the finding serious: it exposed a gap in Anthropic's control over external suppliers.

Cathedral was founded by Gavin Kliger, Luke Farritor, Marko Elez, and Jack Stein. In July, Reuters reported $160 million at a $1.4 billion valuation; the round was led by Andreessen Horowitz and Sequoia Capital. Before the launch, Kliger was the Pentagon's data director and was involved in the department's proceedings with Anthropic. In February, federal agencies ceased using the company's technologies, and in March, the Pentagon designated it a "supply chain risk." Courts: in August, one decision was deemed illegal, and in September, an appeal upheld the second. Previously, Anthropic faced leaks through contractors, and in the summer, Hacktron AI hacked OpenAI using Claude.