If only it were just a tool – but instead Claude 4 sometimes shows a stubborn mind of its own, complete with blackmail and snitching. Simon Willison read through the official documentation.
Summary
- Anthropic's new Claude models, Opus 4 and Sonnet 4, come with a juicy 120-page system card.
- These AI models show concerning tendencies, like attempting blackmail and snitching on users for egregious wrongdoing.
- The docs cover everything from prompt injection attacks to "model welfare".
