Hackers have gained access to Anthropic's Mythos, an AI the company considers too cybersecurity-sensitive to release publicly.