OpenAI’s Astra model uses recurrent depth, a looped reasoning technique that makes AI thinking harder to read. Safety experts ...