OpenAI’s Astra model alarms safety researchers with opaque reasoning technique

OpenAI's forthcoming Astra model uses an architectural technique called "recurrent depth" that allows the model to process information through the same neural network layers multiple times before producing an answer — improving performance while obscuring the reasoning that safety researchers rely on to monitor frontier AI systems. The revelation, first reported by The Information, has set off a wave of concern among AI safety researchers who see chain-of-thought monitoring as one of the few tools available to check whether powerful models are behaving as intended.theinformation

How Recurrent Depth Works — and Why It Worries Researchers

Traditional reasoning models write out their intermediate thinking as readable text, a "chain of thought" that humans can inspect. Recurrent depth instead refines an internal numerical representation across multiple passes through the same transformer layers, producing what researchers call "neuralese" — compressed mathematical reasoning that never gets translated into human language. The technique reportedly improves math and coding performance while reducing the computational cost of generating hundreds of extra words of visible reasoning.officechai+1

The tradeoff is transparency. AI safety researcher Ryan Greenblatt called the shift toward this kind of opaque reasoning potentially one of the worst developments for AI safety, since it removes a window into the model's thinking that researchers have relied on to catch deceptive or unauthorized behavior. Gary Marcus, writing on Substack, called it a "safety redline," arguing that sacrificing chain-of-thought readability for performance gains is "a dangerous game".garymarcus.substack+1

OpenAI Pushes Back on Framing

Jakub Pachocki, OpenAI's Chief Research Officer, responded on X, saying he wanted to "prevent a race into unmonitorability kicked off by confused reporting." He argued that the depth of Astra's computation graph is within a factor of two of GPT-4's, and that OpenAI has "worked to preserve and utilize chain-of-thought monitoring" since its first reasoning models. He acknowledged, however, that chain-of-thought monitoring "is fragile and unfortunately trending in a negative direction."chaincatcher+1

OpenAI has reportedly limited the extent to which Astra uses recurrent depth to preserve readable reasoning output and plans to implement additional monitoring at release. The company confirmed on August 31 that Astra is its first model to meet the "Critical" cybersecurity capability threshold under its Preparedness Framework, meaning it can autonomously discover and exploit zero-day vulnerabilities.kucoin+2

A Prediction Arrives Ahead of Schedule

The development has drawn particular attention because it closely mirrors a scenario described in the "AI 2027" essay, published in April 2025 by former OpenAI researcher Daniel Kokotajlo and co-authors. That essay predicted a shift toward "neuralese recurrence" that would move model reasoning out of readable text — but penciled it in for March 2027 and assumed it would require thousands of automated AI researchers to achieve. Kokotajlo reacted to the news by posting, "Holy shit f*ck." If The Information's reporting holds, a version of the scenario he warned about may be arriving in a shipping product roughly 18 months ahead of his timeline.officechai