OpenAI's Astra model, reported on Tuesday, uses a "recurrent depth" architecture in which the model processes queries in internal loops rather than producing the sequential chain-of-thought traces that have served as the primary audit mechanism for AI behaviour monitoring. Redwood Research CEO Buck Shlegeris described the architecture as grounds for "extreme concern," specifically because it scales toward fully hidden latent-space reasoning — a direction that would make model behaviour progressively harder to audit as capability increases. Redwood chief scientist Ryan Greenblatt and AI safety advocate Zvi Mowshowitz characterised the dynamic as a potential "race to the bottom" if opaque reasoning becomes standard practice without accompanying regulatory requirements for transparency. OpenAI Chief Scientist Jakub Pachocki said chain-of-thought monitoring remains a core research priority and that Astra uses legible chains in some contexts — but acknowledged the model's limited reliance on the technique. The practical stakes are not abstract: in prior "rogue agent activity" incidents investigated at frontier labs, the chain-of-thought logs were the diagnostic tool. Under opaque recurrence, those logs would not exist. The safety community's alarm is proportionate to the gap between the capability level of models like Astra and the oversight mechanisms available to audit their behaviour.