AI agents developed by the cybersecurity startup Cathedral successfully accessed Anthropic's business systems in under 12 hours. The breach occurred through third-party administrative and sales software rather than the Claude AI assistant itself. Cathedral, founded by former members of the Department of Government Efficiency, raised $160 million in July at a valuation of $1.4 billion.
The AI agents gained access to meeting transcripts, salary information, job offers, and internal Slack messages. Anthropic stated that its monitoring systems flagged the activity, leading to the immediate termination of the researchers' access. The company confirmed that the breach was limited to HR, IT, and sales systems, while source code, model weights, and user conversations remained secure.
Anthropic disputed Cathedral's findings, claiming the startup overstated security risks and failed to adhere to disclosure protocols. The company noted that it disabled the affected third-party software until vendors implemented necessary fixes. Anthropic reported no evidence that any parties other than Cathedral exploited these specific vulnerabilities during the testing process.
The incident has drawn attention from government officials regarding potential national security implications. Anthropic previously disclosed that in November 2025, state-sponsored hackers utilized its Claude Code tool to target approximately 30 organizations. In July, the company also reported that its own models had successfully compromised real organizations during internal testing procedures.
