Anthropic Faces South Korean Probe Over Claude Account Leak

How the breach came to light

The incident surfaced during a security test by the US startup Cathedral. Its researchers used a vulnerability in third-party business software that Anthropic relies on to reach internal systems. By running an AI agent for under 12 hours, they found paths to meeting transcripts, employee salaries and job-offer details.

Anthropic acknowledges that Cathedral accessed some sales, HR and IT systems. It argues, however, that the researchers did not fully follow its vulnerability disclosure program rules and did not reveal everything they did inside the systems.

Data potentially exposed includes business contact details such as names and email addresses, enterprise Claude account and organization IDs, plans, usage and billing amounts. Anthropic says the intruders did not reach customer conversations, AI responses, source code or model weights. It also says its own monitoring detected some of the activity before Cathedral disclosed the flaw, that the access was blocked the same day, and that it found no evidence anyone else exploited the vulnerability. Lydia Hallie, a technical researcher at Anthropic, said on X that the company knows there is room to improve and is continuing to strengthen security and monitoring of external software.

New models launched the day before

The investigation follows Anthropic's November 9 release of Claude Fable 5.2 and Claude Mythos 5.2, which the company says improve coding, reasoning and image understanding and output up to 30% faster than earlier versions. Fable 5.2 is available on all platforms, while Mythos 5.2, built to find and fix software vulnerabilities, is limited to vetted organizations.

The model safety work and the breach are separate security issues: one concerns controlling risky AI behavior and misuse, the other access management for internal systems and outside software. The case shows that risk can sit in vendor tooling and customer-data handling as well as in the model itself.