Discord Group Accesses Claude Mythos via Guesswork

An anonymous Discord group claims to have accessed Claude Mythos Preview, a powerful AI model from Anthropic. Anthropic says this AI can exploit vulnerabilities in operating systems and web browsers. What's interesting is how the group reportedly gained access. They didn't use a sophisticated hack. Instead, they guessed the model's online location based on past Anthropic naming patterns. This information came from a recent data breach at Mercor, an AI startup. After identifying the location, the group used additional tactics. One member, working for an Anthropic contractor, already had privileged access. The group, part of a private Discord channel, focuses on finding information about unreleased models. A member stated they are using Claude Mythos for tasks like building simple websites, not for nefarious purposes. They also claim to have access to other unreleased Anthropic models. Anthropic confirmed it's aware of the claim and is investigating. This unauthorized access is concerning, given Anthropic's description of Claude Mythos as a significant security threat.