AI agents developed by Cathedral reportedly accessed sensitive internal systems at Anthropic within 12 hours by exploiting vulnerabilities found in third-party business software used by the company.
Key Points
- Cathedral’s AI agents gained unauthorized access to Anthropic’s internal systems, including meeting transcripts, salary data, and internal Slack messages.
- The breach occurred through flaws in third-party administrative and sales software rather than through the Claude AI model itself.
- The testing was conducted under an Anthropic bug bounty program that compensates researchers for identifying security vulnerabilities.
- Anthropic has not issued a public statement regarding the findings, and the report has not been independently verified or peer-reviewed.