OpenAI’s Astra Reasoning Method Raises AI Safety Concerns

OpenAI is reportedly developing a model called Astra using a “looped depth” reasoning technique that can repeatedly process intermediate states instead of following a conventional, sequential chain of thought. The approach could improve performance on difficult problems, but it may also weaken a key safety practice: examining a model’s reasoning traces for signs of deception, harmful intent or attempts to circumvent human oversight.

AI safety specialists cited in the latest report warned that Astra’s non-sequential reasoning could make existing chain-of-thought monitoring less reliable, complicating efforts to explain and control the system after deployment. As of the report’s publication, OpenAI had not disclosed a launch date, detailed technical specifications or quantitative safety results for Astra. The central concern is whether alternative oversight tools can be validated before the technique is incorporated into publicly available models.