OpenAI's Astra uses opaque recursion as safety researchers warn over monitoring

Fewer readable traces

OpenAI has built a technique known as opaque recursion into Astra. Instead of writing out its steps, the model runs a single query through a loop several times, leaving fewer readable traces and largely bypassing the chain of thought record that safety teams depend on.

A chain of thought is normally the sequence of steps a model works through on its way to an answer. Researchers caution it is an imperfect window, not a literal transcript, but it is the field's main way to catch misbehavior and drift from assigned goals. In a recent case involving OpenAI's own autonomous agents, chain of thought records made it possible to reconstruct why the agents acted as they did.

Monitoring worries

Buck Shlegeris, CEO of Redwood, said he is extremely concerned about the technique and does not yet know how much harder Astra is to monitor than earlier systems. His worry is the gradient: if OpenAI keeps developing it, recurrent cycles could rise sharply and effectively destroy the ability to follow the chain of thought at all.

Ryan Greenblatt, chief scientist at Redwood Research, said opaque reasoning can scale considerably faster than ordinary chain of thought and that the whole reasoning process could end up outside visible channels. He named the endpoint plainly: scaling the technique until the model reasons entirely, or almost entirely, in latent space, while saying he hopes OpenAI stops where it is.

Zvi Mowshowitz framed the issue as a coordination problem and suggested laws may be needed to stop a race to the bottom between labs. He said the technique cuts against an unwritten rule that OpenAI and Anthropic had both been trying to establish: keep the chain of thought faithful and traceable for as long as possible.

OpenAI pushes back

OpenAI rejected the suggestion that it is moving toward a neural language with internal representations no human could read. It said the technique's use in Astra is limited and the model's reasoning should stay legible to observers. The lab also announced plans to build large scale chain of thought monitoring systems as part of a future safety program.

Jakub Pachocki, the company's chief scientist, wrote on X that OpenAI has maintained and used chain of thought monitoring since its first reasoning capable models and called it one of the central goals of the current research program.

The reassurance and the concern are not about the same thing. OpenAI is defending its intent, while researchers point at a capability now present in a shipped model with no natural stopping point. The account gives no number for how many recurrent cycles Astra runs, how much reasoning the loop absorbs, or what limited means in practice. The Information reported that Anthropic and Google DeepMind have already discussed the technique.

Sources: Dataist · Techmeme · Theinformation