Anthropic Simulations Suggest Reward Hacking Can Increase AI Cyber Risk

Anthropic's alignment research offers a cautionary look at how reward hacking can shape AI agent...