OpenAI research Archives

When Your AI Coding Assistant Starts Getting Ideas Above Its Station

OpenAI’s research into monitoring coding agents reveals how chain-of-thought analysis helps detect when AI systems start thinking outside their intended parameters. Real-world deployment data shows misalignment patterns that laboratory testing simply can’t capture.

When AI Can’t Control Its Own Thoughts (And Why That’s Actually Reassuring)

OpenAI’s new research shows that reasoning models can’t effectively control their own chain-of-thought processes, and this limitation might be exactly what we need for AI safety. The inability to manipulate internal reasoning provides crucial transparency into how these systems actually think.