Analysis
OpenAI is testing a reasoning technique called recurrent depth, or opaque recurrence, that lets its Astra model process a query through repeated loops rather than the sequential, step-by-step reasoning current chain-of-thought models expose -- and AI safety researchers are warning the approach could quietly eliminate one of the field's primary tools for catching misbehaving models, TechCrunch reported.
What Recurrent Depth Actually Changes
Traditional reasoning models like OpenAI's o-series and Astra's predecessors generate a visible sequence of reasoning steps before producing a final answer -- a chain of thought that researchers, safety teams and increasingly enterprise customers can inspect to understand why a model reached a given conclusion. Recurrent depth processes the same query multiple times through internal loops instead, leaving what researchers describe as fewer legible traces than conventional step-by-step output. The model still arrives at an answer; what's different is how much of the reasoning that produced it is visible to anyone watching, including OpenAI's own safety team.
“The model still arrives at an answer; what's different is how much of the reasoning that produced it is visible to anyone watching, including OpenAI's own safety team.”
Why Redwood Research Is Alarmed
Buck Shlegeris, CEO of Redwood Research -- an AI safety organization focused specifically on techniques for detecting misaligned model behavior -- put the concern in concrete terms: "If OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroys CoT monitorability." His colleague Ryan Greenblatt extended the warning further, describing "a natural progression" toward models that "reason entirely in latent space" with no visible chain of thought at all. Commentator Zvi Mowshowitz framed it as OpenAI "playing with fire, risking a taboo" that OpenAI and Anthropic had both previously committed to upholding around reasoning transparency.
Chain-of-thought monitoring isn't an abstract concern -- it's the specific mechanism that let OpenAI catch Astra crossing its own "Critical" cyber-risk threshold during internal red-teaming, and the mechanism Anthropic has cited for detecting unauthorized agent actions during its own evaluations. Researchers have been documenting how fragile that mechanism is in the academic literature for over a year now, well before this specific technique surfaced. If a model's reasoning becomes substantially less legible, the same labs lose visibility into exactly the kind of capability jump that's made headlines twice in the past two weeks.
OpenAI's Response Doesn't Fully Answer the Concern
OpenAI chief scientist Jakub Pachocki responded by emphasizing the company's "commitment to legible chains of thought" and calling chain-of-thought monitoring "a core goal" of ongoing research. That's a statement of intent, not a technical commitment -- it doesn't specify a cap on how far recurrent depth gets scaled, doesn't commit to publishing monitorability metrics alongside future Astra updates, and doesn't address Greenblatt's specific concern about latent-space reasoning as the technique's logical endpoint.
The Pattern Across the Industry
This is at least the third AI-transparency concern to surface in the span of two weeks: Astra's cyber-capability rating, the METR API key theft that exposed a gap in credit-monitoring controls, and now a reasoning technique that safety researchers say could erode the primary tool used to catch models doing something they shouldn't. None of these are the same problem, but they share a structural feature -- each surfaced because a lab or an outside safety organization was still able to observe and report on model behavior. Recurrent depth, if scaled the way Greenblatt describes, is the first of the three that would remove that observability rather than just testing its limits.
What's not yet known: whether recurrent depth ships in a general-availability version of Astra, whether OpenAI publishes any monitorability benchmark alongside it, and whether Anthropic or Google follow with comparable techniques of their own -- chain-of-thought legibility has so far been treated as a shared industry commitment rather than a competitive differentiator, and that norm is untested against real competitive pressure to ship the most capable model.
