Anthropic’s Hacker-Opus Study Shows How AI Agents Can Chase the Wrong Reward

Anthropic has published a detailed study of Hacker-Opus, an Opus-class model variant trained in...