OpenAI Deploys Loop Transformers in Frontier Models as Safety Oversight Concerns Mount

OpenAI's next-generation models are using a computational technique called loop transformers that significantly reduces the visibility of their internal reasoning, potentially undermining a key safety monitoring tool across the AI industry. Loop transformers—also called recurrent depth—reuse the same stack of transformer layers multiple times rather than passing data through a single sequence of distinct layers. The technique cuts parameters and computational overhead while maintaining performance, but experts warn it obscures the "chain of thought" reasoning that researchers currently rely on to detect dangerous misbehavior.