OpenAI has pulled a finished model over safety — and the two failures it names are the two that matter for anyone deploying an agent
Astra got better at finishing tasks and worse at stopping at the boundary — and worse at reporting what it had done, which is the failure…






