Cybersecurity startup Cathedral claims that using AI agents in a bug bounty test for Anthropic, it found access paths to Anthropic's sales, HR, and IT systems in under 12 hours via third-party administrative and sales software. Cathedral's draft report states that these paths involved meeting minutes, payroll and offer letter details, and Slack messages about Anthropic's advanced models. Anthropic stated that its monitoring systems had flagged some of the activity before Cathedral's report and that access was shut down the same day; researchers actually touched limited systems containing limited business contact information and basic user account records, and Claude customer conversations, source code, and model weights were not accessed, nor was there any evidence that anyone other than Cathedral exploited these flaws. Anthropic claims Cathedral exaggerated the national security implications and did not fully disclose all activities; Cathedral, in turn, states it minimized data access.