Startup accesses Anthropic data

The Washington Post reported that a young cybersecurity startup named Cathedral accessed internal data from Anthropic, a company that develops the AI model Claude. The breach was discovered through a draft of Cathedral's report, which detailed how its AI agents found vulnerabilities in Anthropic's systems within less than twelve hours.

Vulnerabilities in third-party software

Cathedral did not directly target Anthropic's systems or models. Instead, its researchers exploited vulnerabilities in third-party software that Anthropic uses for administrative and commercial tasks. Through these vulnerabilities, Cathedral researchers claim they accessed meeting transcripts, salary information, job offers, and Slack messages discussing the company's advanced models.

Bug bounty program and expert review

Cathedral explains that it conducted these tests as part of Anthropic's bug bounty program and shared its findings with the company and relevant publishers at the beginning of October. Alex Stamos, a former director of the Stanford Internet Observatory, reviewed the results and noted that the issue highlights a weakness in how Anthropic manages its external vendors, though he acknowledged that connecting third-party applications to internal systems is a common problem.

Anthropic's response

Anthropic acknowledges that a "limited set" of commercial, human resources, and IT systems were accessed, but it claims the scope was much more restricted. The company states that the exposed data included only professional contact information for clients and partners and basic user account details, while conversations with Claude, the code, and model weights remained secure.

Cathedral's claims and national security concerns

Cathedral argues that the incident is a warning for all American companies that rely on ordinary software for their most sensitive infrastructures. A high-ranking Trump administration official raised concerns about national security, suggesting that if a small startup could do this, the Chinese may not be far behind.

Political and corporate tensions

Anthropic has tense relations with Washington after the Pentagon excluded it from federal systems this year due to its refusal to allow its technology to be used for autonomous weapons or mass surveillance. Cathedral, on the other hand, has close ties to the Trump administration, with co-founders linked to the DOGE group and existing contracts with the federal government.

Broader implications for AI security

The incident has sparked debate beyond the technical details. Clement Delangue of Hugging Face argues that concentrating AI in closed environments does not ensure security. Shyam Sankar of Palantir suggests that cutting-edge AI should be treated as a dual-use national power, requiring both defensive and offensive capabilities from American actors.