The key facts

  • Cathedral, a startup founded by former DOGE employees, gained access to Anthropic's internal systems in less than 12 hours
  • Meeting transcripts, salary data and Slack correspondence about AI models were accessible
  • Anthropic confirmed access to a 'limited set' of systems and said it had closed the vulnerabilities and informed federal authorities

How a startup got into Anthropic's systems

Cathedral, a startup that has existed for only a few months and was founded by former participants of the US Department of Government Efficiency (DOGE), gained access to sensitive internal data of Anthropic — the developer of the Claude chatbot — in less than 12 hours. This was reported by The Washington Post citing a draft report that the Anadolu agency retold.

What turned out to be accessible and how it was organized

The researchers used AI agents to find paths into Anthropic's internal systems. As a result, they gained access to meeting transcripts, salary and job offer information, as well as Slack correspondence in which advanced AI models were discussed. The testing was carried out under a program that allows external cybersecurity researchers to look for vulnerabilities in the company's systems. Notably, according to an Anthropic representative, Cathedral spent several days unsuccessfully trying to hack the company's own software before switching to third-party business tools — and it was through them that the way inside was found.

The reaction of Anthropic and US authorities

Anthropic confirmed that Cathedral gained access to a 'limited set' of its sales, HR and IT systems. The company said it had closed the vulnerable access paths and informed federal officials about the incident. According to the publication, a senior Trump administration official said the report circulated within the government and raised concern about the speed with which the startup gained access to sensitive systems. The official, according to The Washington Post, warned that these findings raise the question of whether foreign intelligence services could exploit similar vulnerabilities.