Unauthorized Access to Claude Mythos

Anthropic is investigating reports of unauthorized access to Claude Mythos Preview in a third-party vendor environment. According to Bloomberg, a small number of Discord users entered a restricted test environment, and Anthropic stated that there was no breach of core systems. The model was first deployed to a few approved enterprises in late February, and publicly introduced on April 7.

The access route was reportedly a single anonymous contractor employee account and the use of common internet investigation tools. Claims of ShinyHunters involvement quickly lost credibility as apparently AI-generated fake screenshots spread, and researchers also denied the claims.

• Mythos is a high-risk model aimed at vulnerability detection and automated patching.
• Project Glasswing is provided on a limited basis to only 40 organizations, including Amazon, Google, Microsoft, Apple, and Cisco.
• This incident showed that third-party access controls could be the biggest vulnerability of high-performance AI models.