OpenAI's Astra Model Raises AI Safety Red Flags

OpenAI's forthcoming Astra model uses a reasoning method called "recurrent depth," according to a report from The Information published Tuesday. Also known as "opaque recurrence," the technique lets Astra process information outside the step-by-step sequence that most reasoning models rely on. That sequential structure is what normally produces a chain of thought, a written record of a model's reasoning that researchers use to catch problems before they surface in output. Recurrent depth breaks from that pattern, and the report says Astra's use of it is limited for now.

Limited or not, the disclosure landed hard in AI safety circles. Chain of thought monitoring has become one of the main tools labs use to check whether a model is reasoning honestly or hiding something in its process. Anthropic and OpenAI have both treated preserving that visibility as a shared priority, even a kind of informal red line between competitors. A technique that makes reasoning harder to read threatens that arrangement, and safety researchers wasted no time saying so once the report circulated.