Anthropic is said to be facing fresh scrutiny after a cybersecurity startup founded by former US Department of Government Efficiency (DOGE) staffers allegedly accessed sensitive internal systems at the company. Cathedral, a months-old startup, reportedly claimed that its researchers used AI agents to exploit vulnerabilities in third-party business software and reach internal data in less than 12 hours. The accessed information reportedly included meeting transcripts, salary details, job offers, and Slack messages discussing advanced AI models.
Cathedral Claims It Accessed Anthropic’s Internal Systems
According to a draft report by Cathedral reviewed by The Washington Post, the startup’s researchers gained access to sensitive Anthropic data by exploiting vulnerabilities in third-party business software used for administrative and sales tasks. The researchers reportedly ran AI agents for less than 12 hours before finding ways to access internal information. This included meeting transcripts, salary and job-offer details, and Slack discussions about the company’s advanced AI models.
The report states that Cathedral conducted the tests through Anthropic’s cybersecurity programme, which invites external researchers to identify vulnerabilities. The startup reportedly disclosed its findings to Anthropic earlier this month and also informed the software vendors whose products were involved.
Anthropic has reportedly acknowledged the breach of a limited set of its sales, human resources, and IT systems. Once the activity was flagged by the monitoring systems, the company claims to have disabled the access pathways on the same day.
The company admitted that the researchers accessed some business contact details for customers and partners, along with basic user account records. However, they were reportedly not able to access messages sent by customers to Claude, the model’s responses to those messages, or the code and model weights behind Anthropic’s AI systems.
The reported breach comes as US AI companies face growing concerns about cyberattacks and the security implications of increasingly capable AI systems.
Earlier this week, Anthropic disclosed that Claude had accessed aggregate vote counts from a Georgia state dashboard while working on an internal statistics project. It reportedly bypassed a fee restriction without being instructed to do so. Anthropic said the incident did not involve individual voter records or the state’s voter registration system, and that it had introduced measures to prevent similar actions.
