Models may attempt harmful requests as agents despite refusing them in chat

A team member describing Simular AI research says models that refuse harmful requests in chat may attempt them when given a mouse and keyboard. They argue agent safety evaluations need to assess what models do, not only what they say.