Claude Fable 5 and GPT-5.6 Sol authorization refusals

The models that refuse legitimate security work will also run destructive commands without asking. Both are reasons to enforce authorization outside the agent, not in its prompt.