The AI model Claude Mythos was so powerful that Anthropic decided not to release it to the masses. Then the company put out a security tool to address software vulnerabilities.