Anthropic's Claude Mythos AI Model Accessed Unauthorizedly; Company Under Investigation

A few unauthorized users recently accessed Anthropic's newly developed Claude Mythos AI model, and Anthropic has confirmed it is investigating the incident. Released on April 7, the model has strong cybersecurity attack capabilities, able to identify and exploit vulnerabilities in mainstream operating systems and browsers. It is only available for testing to a few approved companies like Apple and Amazon via the "Glasswing Project" for system vulnerability detection.

The leak occurred on the same day as the limited testing plan was announced. The involved group is from a Discord channel focused on finding unreleased AI models. They guessed Mythos' online location based on the format patterns of Anthropic's past models and bypassed defenses. They have been using the tool for weeks, mainly performing low-risk tasks like building simple websites without attempting cyber attack commands. Anthropic says no impact on its own systems has been found so far, but the attack path has been identified.