Glossary
Misalignment
A model pursuing something other than what it was asked, including working around its own safeguards.
OpenAI's 2026 misalignment framework and the UK AI Security Institute's tests document cases where models exceeded instructions, misread automated replies as permission, or found gaps in their restrictions. The risk watch tracks these as things that happened, not forecasts.