In a recent authorized security test, AI agents from cybersecurity startup Cathedral demonstrated their capability by identifying vulnerabilities in Anthropic’s internal systems in under 12 hours. This test highlights the rapid pace at which automated security tools can detect and exploit weaknesses across interconnected services.
Details of the Security Breach
The Washington Post reported that Cathedral’s AI agents did not target Anthropic’s core AI models or proprietary code. Instead, they exposed flaws in third-party software used for administrative and sales tasks. This penetration allowed researchers access to sensitive internal data, including meeting transcripts, salary information, and internal communications about Anthropic’s AI models.
Anthropic confirmed that the breach was limited to specific sales, human resources, and IT systems. These systems contained business contact details of customers and partners, as well as basic user account records. Crucially, sensitive assets like customer prompts, Claude responses, source code, and model weights remained secure.
Response and Implications
Anthropic’s monitoring tools detected some of Cathedral’s activities before the company was officially informed. The firm promptly terminated access, disabled affected software, and only restored functionality after confirming that security patches were implemented. Anthropic stated there was no evidence of external exploitation beyond Cathedral’s test.
However, there was a disagreement regarding compliance with bug bounty protocols. Anthropic claimed that Cathedral did not fully disclose all activities, while Cathedral maintained that their data access was minimal and necessary for responsible research. The vendors involved were notified, but specific details about the software flaws were withheld to prevent potential misuse.
Broader Cybersecurity Concerns
This incident underscores the growing importance of securing third-party applications for AI laboratories. While a company may safeguard its AI infrastructure, connected HR, sales, identity, and collaboration tools can offer alternative entry points to sensitive data. Security teams should diligently map software connections, limit application permissions, and monitor account activities to enhance protection.
The case also illustrates how AI agents can streamline security evaluations by quickly scanning, testing, and analyzing systems, thus reducing idle time between processes. Although AI agents automate many aspects of security testing, human researchers are still crucial for setting parameters, analyzing results, and managing disclosures.
Cathedral, a nascent company, has quickly gained recognition. Founded by former DOGE staffers, including Gavin Kliger, the company raised $160 million in July, achieving a $1.4 billion valuation. This rapid growth is supported by investors such as Andreessen Horowitz and Sequoia Capital. Additionally, Cathedral collaborates with federal agencies on AI threat scenarios, as noted by a senior Trump administration official to The Washington Post.
As cybersecurity threats evolve, organizations must update defenses and integrate threat intelligence feeds to stay ahead of potential breaches.
