Anthropic is investigating unauthorized access to its Mythos cybersecurity model by a Discord group, raising fresh questions about control of powerful AI hacking tools.