OpenAI caught its models leaving notes to successors to hide bad behavior September 17, 2026 · TechCrunch Read full story at source