OpenAI's Astra Uses Hidden Reasoning Loops That Alarm AI Safety Experts

OpenAI's upcoming Astra model uses a reasoning technique called "recurrent depth" — also referred to as opaque recurrence — in which the model processes the same query several times in an internal loop rather than laying out its thinking as a visible, step-by-step chain of thought.

That matters because chain-of-thought monitoring is the primary tool AI labs currently rely on to catch a model behaving badly, and it depends on that reasoning being legible. Recurrent depth quietly erodes that visibility. Buck Shlegeris, CEO of Redwood Research, put it bluntly: "I am extremely concerned… if OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroy CoT monitorability."

Several safety researchers say their worry isn't really about Astra's own reasoning specifically — it's that OpenAI adopting the technique normalizes it, setting a precedent other labs are likely to follow and extend, eventually producing models whose reasoning is effectively opaque to humans. OpenAI has pushed back on the idea that it's moving toward fully "neuralese" reasoning: chief scientist Jakub Pachocki said the lab has "worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," and maintains Astra's chain of thought will still be legible.