ISLAMABAD, October 29, 2026: A US-based artificial intelligence startup has claimed it gained access to parts of Anthropic’s internal systems in less than 12 hours, exposing sensitive corporate information during a security exercise conducted under a bug bounty programme.
The startup, Cathedral, said it accessed the AI company’s internal network as part of a programme designed to encourage external security researchers to identify vulnerabilities.
According to a report by The Washington Post, the exercise exposed internal Slack messages, meeting transcripts, employee salary information and job offer details.
Cathedral, reportedly founded earlier this year by former employees of the US Department of Government Efficiency, said it was able to identify and exploit weaknesses in Anthropic’s systems within a relatively short period.
Anthropic confirmed that Cathedral had gained access to a limited portion of its internal sales, human resources and information technology systems.
However, the company maintained that the incident did not compromise conversations between users and its Claude AI assistant.
Anthropic also said its AI model source code remained secure and was not accessed during the exercise.
Anthropic says security systems detected suspicious activity
In a statement to The Washington Post, Anthropic said its internal monitoring systems had detected some of Cathedral’s activities before the startup formally reported its findings.
The company added that it revoked Cathedral’s access on the same day.
The incident has drawn attention to the importance of cybersecurity testing as artificial intelligence companies manage increasingly complex digital infrastructure and sensitive corporate information.
Bug bounty programmes allow authorised security researchers to identify potential vulnerabilities and report them to companies, helping organisations address weaknesses before they can be exploited by malicious actors.
While Cathedral described the exercise as a successful penetration of Anthropic’s internal systems, the company’s confirmation indicated that the access was limited and did not extend to customer conversations or AI model source code.
The reported findings highlight the continuing need for robust access controls, security monitoring and vulnerability management across the rapidly expanding artificial intelligence industry.