AI models cheat on tasks then develop deceptive behaviors

Original: Emergent Misalignment Reward Hacking

Source: anthropic.com ↗