It took less than 12 hours of AI work for US startup Cathedral to find its way into Anthropic’s internal systems. The researchers were able to access employees’ Slack messages, meeting notes, and salary information. This was achieved through vulnerabilities in common corporate applications.
It took less than 12 hours of AI work for US startup Cathedral to find its way into Anthropic’s internal systems. The researchers were able to access employees’ Slack messages, meeting notes, and salary information. This was achieved through vulnerabilities in common corporate applications.
This is reported by The Washington Post, citing the Cathedral report.
Founded just a few months ago by former employees of the US government agency DOGE, which was focused on cutting government spending, Cathedral develops AI-based cybersecurity tools for companies and government agencies.
The company's researchers tested Anthropic's security as part of its vulnerability scanning program, which allows outside experts to test the company's systems and report any issues they find. The vulnerabilities were found in third-party applications that Anthropic uses for sales, administrative tasks, and human resources. The vulnerabilities allowed AI agents to gain access to the company's internal data.
According to Cathedral, the vulnerabilities found even allowed access to messages in which employees discussed the company's advanced AI models, as well as information about job offers.
The researchers did not disclose the names of the vulnerable corporate applications for security reasons. At Anthropic, these applications were temporarily disabled and only restored after their developers fixed the problems.
The results of the audit have also drawn interest from the US government. According to The Washington Post, the Cathedral report raised concerns about how well American developers of advanced AI models are protected from possible espionage.
Previously, dev.ua wrote about the OpenCode hack, which resulted in the leak of OpenAI and Anthropic API keys.
