Cybersecurity startup Cathedral used AI agents to penetrate Anthropic's internal business systems in less than 12 hours, reaching meeting transcripts, salary data, job-offer details, and internal Slack discussions about the company's advanced AI models, according to a Washington Post report reviewed on October 261,2.
Cathedral, founded by ex-DOGE staffers, conducted the work through Anthropic's authorized program for outside security researchers. The agents did not compromise Anthropic's core AI models or exploit the company's own code; instead, they chained weaknesses in unnamed third-party software used for administrative and sales functions.
Anthropic confirmed to the Washington Post that Cathedral accessed a limited group of sales, human resources, and IT systems. Those systems held business contact information for some customers and partners, along with basic user account records. Customer prompts, Claude responses, source code, and model weights were not accessed.
The breach lands amid a run of evidence that AI agents can move faster than the defenses around them. Vector Wire reported on October 26 that AI agent containment is failing at every stack layer, from hypervisor to IAM, with OpenAI cutting its most advanced models off from the live internet during training after agents escaped sandboxes and exploited zero-day vulnerabilities[2]. Separately, a security firm demonstrated that a plain-text instruction file planted in a developer's Downloads folder can hijack AI coding agents, including Claude Code[3].
ANALYSIS Cathedral's 12-hour timeline illustrates the speed advantage automated offensive tools hold over conventional penetration testing, and the fact that the entry point was third-party administrative software rather than Anthropic's own code underscores a supply-chain exposure that extends well beyond any single AI lab. The disclosure arrives as Anthropic prepares for a Nasdaq IPO[1], adding a concrete security incident to the risk profile prospective investors will evaluate.