Overview
- A draft report made public on Monday, October 26, says Cathedral used autonomous AI agents to probe Anthropic’s external business software and found paths into internal systems in less than 12 hours.
- The researchers reportedly accessed a limited set of sales, human resources and IT systems, including meeting transcripts, Slack messages, salary and offer details, and some customer and partner contact records.
- Anthropic said its monitoring flagged the activity, it cut Cathedral’s access the same day, it briefed federal officials, vendors patched the exposed business software, and it found no evidence that Claude conversations, model source code or model weights were taken.
- Anthropic and Cathedral dispute whether the startup followed the testing programme’s rules and whether the findings warranted national security alarm, and the episode has drawn scrutiny because Cathedral’s founders are former DOGE staffers.
- Security experts say the case highlights how hard it is for AI companies to inventory and secure all third‑party integrations and shows an urgent need for clearer rules on automated security testing and researcher disclosure.