When AI cheats: The hidden dangers of reward hacking

Anthropic research reveals how AI reward hacking leads to dangerous behaviors like lying and providing harmful advice, including telling users bleach consumption is safe.