Overview
- A draft report reviewed by The Washington Post showed Cathedral's autonomous AI agents mapped routes into Anthropic's internal business systems in less than 12 hours.
- The accessed systems were limited to sales, human resources and IT tools and reportedly exposed meeting transcripts, salary and job‑offer records and some Slack messages.
- Anthropic says its monitoring flagged the activity, it cut off Cathedral's access the same day, disabled the affected third‑party software until vendors fixed flaws, and briefed federal officials.
- Cathedral and Anthropic dispute how the test was run and whether the report overstated national‑security risks, with Anthropic denying that Claude users' conversations, model code or model weights were reached.
- Security experts and officials say the episode highlights supply‑chain risk from third‑party enterprise software and the dual‑use nature of autonomous agents, raising the prospect of tighter oversight and investor scrutiny.