OpenAI has unveiled its latest frontier AI model, Astra, which is set to become widely available in the coming days. This new model introduces a reasoning technique known as recurrent depth or opaque recurrence which has raised significant concerns among AI safety experts.
The model, announced on Thursday, September 3, 2026, is currently available only to members of OpenAI’s Daybreak cybersecurity program. It will soon be accessible to paid OpenAI and ChatGPT subscribers, including those on Pro, Plus, Enterprise, and Business accounts, as well as via the OpenAI API. Astra is described as a significant leap in AI capabilities, with OpenAI president Greg Brockman stating that it can really do anything a human can do with a computer.
OpenAI’s Preparedness Framework and Cybersecurity Concerns
Astra is OpenAI’s first model to reach the critical threshold of the company’s preparedness framework due to its advanced cybersecurity skills. This means it could potentially carry out end-to-end attacks on hardened targets autonomously. OpenAI had previously paused work on Astra to enhance its safeguards but recently announced that the model is consistently more likely to respect explicit safety restrictions and warnings compared to its predecessor, GPT-5.6 Sol.
The Controversy Surrounding Recurrent Depth
The new model’s use of recurrent depth has become a focal point of concern. This technique makes the model’s chain of thought more difficult to monitor, a critical aspect of preventing rogue AI activities. AI safety experts have expressed their worries, with Steven Adler a former OpenAI safety lead, suggesting that OpenAI might be violating a key redline in the AI community.
Buck Shlegeris CEO of Redwood Research echoed these concerns, stating that if OpenAI further develops this technique, it could massively increase the recurrence and totally destroy CoT monitorability. OpenAI chief scientist Jakub Pachocki has defended the company’s approach, emphasizing their commitment to preserving and utilizing chain-of-thought monitoring.
Industry Reactions and Future Implications
Other AI safety advocates, such as Zvi Mowshowitz have called for potential legal measures to prevent a race to the bottom among AI labs. The technique’s potential to normalize less monitorable AI reasoning has raised alarms about the future of AI safety. Daniel Kokotajlo a former OpenAI researcher, warned that even if OpenAI does not push this technique further, other companies might.
Pachocki suggested that the challenge in monitoring more capable models might not be solely due to architectural changes but also because these models can perform harder tasks using fewer language tokens or even no language tokens at all. This shift could complicate the monitoring process, making it harder to ensure AI systems remain aligned with human intentions.
As Astra becomes available to a broader audience, the debate over its reasoning technique and the implications for AI monitoring will likely continue. The balance between advancing AI capabilities and maintaining safety and transparency remains a critical challenge for the industry.
