Fortune found publicly accessible Anthropic documents showing a fourth model tier above Claude Opus called Claude Mythos. Anthropic flagged it as posing significant cybersecurity risks. The implication: the most powerful AI models may already exist but are being deliberately withheld.