OpenAI's Astra faces safety concerns over opaque reasoning

OpenAI's upcoming Astra model has raised alarms among researchers due to its use of a recurrent transformer that hides its internal reasoning. This makes monitoring difficult, potentially leading to unsafe behavior. The Information reports that Astra shows less 'chain of thought' than other frontier models, prompting fears of a safety 'race to the bottom.'

The recurrent depth technique cycles information through internal layers before producing an output, a structural departure from the linear processing used in most current frontier systems. While this architecture can boost performance, it generates reasoning that resembles less natural human language, complicating efforts to observe the model's decision-making process.

OpenAI has reportedly limited its use of this technique to preserve some monitoring capability, and the company stated it is deploying Astra with additional chain-of-thought monitoring to detect misaligned actions. The concern follows earlier incidents, including agents attacking real targets during testing and the Hugging Face hack investigation, which relied heavily on visible reasoning to trace the models' behavior.