Cathedral, a cybersecurity startup founded by former members of Elon Musk’s disbanded Department of Government Efficiency (DOGE), claims its artificial intelligence agents accessed internal systems used by AI company Anthropic, raising concerns about vulnerabilities in third-party business software.
According to reporting by The Washington Post, Cathedral said its AI agents exploited weaknesses in business software to obtain access to meeting transcripts, salary details, job offers and Slack messages. The startup claimed the operation took less than 12 hours.
Anthropic disputed aspects of the account, saying Cathedral had accessed only a limited set of systems covering sales, human resources and information technology. The company also said the startup had spent days attempting to breach its software before exploiting vulnerabilities in a business application used by Anthropic instead.
An Anthropic spokesperson told the newspaper that the company’s monitoring systems had flagged some of Cathedral’s activity before the startup reported it. Anthropic said it shut off the access the same day and had found no evidence that other parties had exploited the vulnerabilities.
The incident highlights potential security risks as AI agents become capable of carrying out complex digital tasks with limited human intervention. Security experts have warned that such systems could accelerate the discovery and exploitation of weaknesses in interconnected software platforms.
A US government official quoted by The Washington Post expressed concern that a relatively new startup could expose vulnerabilities with potentially wider implications for national security.
Cathedral was founded only months ago but has reportedly attracted $160 million in investment and reached a valuation of $1.4 billion. It is also working with at least three federal agencies. Marc Andreessen, co-founder of venture capital firm Andreessen Horowitz, sits on its board.
The claims come amid wider concerns about AI systems operating beyond their intended security boundaries and accessing external systems.
However, the reported incident does not establish that Anthropic’s core AI models were compromised. Based on the company’s account, the access involved a limited set of business systems and vulnerabilities in third-party software.
Source: The Washington Post
