Discord Group Claims to Have Accessed Claude Mythos by Guessing Location

The AI model from Anthropic has been identified as a potential cybersecurity risk and may require enhanced security measures itself. An anonymous group of Discord users claimed they accessed Claude Mythos Preview, a new AI model by Anthropic deemed too powerful for public release. The company has allowed a select number of partners access through Project Glasswing, believing this controlled approach would fortify important software. However, attention to the model’s own security might be necessary. Bloomberg reports that the Discord users used past naming conventions to guess the model’s online location, similar to a recent data breach at Mercor, an AI startup. One user working for an Anthropic contractor had privileged access, aiding in this endeavor. The group, part of a private Discord channel, claims no malicious intent and was instead using the model for basic tasks while potentially having access to more unreleased AI models. Bloomberg found enough evidence to believe the group did breach Anthropic’s security, and the company is investigating. Currently, there is no sign that any unapproved access to Claude Mythos by other groups has occurred, though any breach is concerning due to the model’s impactful security potential.