According to TechCrunch, citing The Information, OpenAI’s upcoming Astra model will employ a reasoning technique called "recursive depth." This approach may reduce visible traces during the model’s reasoning process, making it harder for outsiders to determine why the model reached a particular conclusion through chain-of-thought logs.
Security researchers have issued a warning
In common reasoning models, chain-of-thought typically demonstrates step-by-step how the model processes a problem. Although this record does not equate to the model’s full internal process, it remains an important tool for observing deviations from goals, anomalous decisions, or potential失控 behaviors.
Buck Shlegeris, CEO of Redwood Research, expressed serious concern about Astra’s use of such technology. He believes that if OpenAI further expands the use of these methods, the traceability of the model’s reasoning chains could significantly decrease.
Zvi Mowshowitz, a commentator long focused on AI safety, also stated that if laboratories compete over model capabilities, regulators may eventually need to intervene to prevent the industry from continuously lowering its safety standards.
Reduce visible traces through circular reasoning
The report mentions that so-called "opaque recursion" refers to models no longer completing reasoning primarily through linear steps, but instead iteratively processing the same problem. This approach may reduce the visibility of intermediate processes, effectively bypassing the observable pathways provided by traditional chain-of-thought logging.
This is also the part that security researchers are most concerned about, because as models increasingly perform reasoning within invisible internal spaces, it becomes harder for researchers to determine whether they are misleading, circumventing constraints, or exhibiting other unintended behaviors.
Ryan Greenblatt, Chief Scientist at Redwood Research, said the greater risk is that this kind of opaque reasoning may scale more easily than traditional chain-of-thought reasoning, ultimately moving the vast majority of reasoning processes outside visible channels.
OpenAI emphasizes maintaining monitoring capabilities.
However, Astra's current use of this technology is reportedly still limited. The report notes that the model's chain of thought is expected to remain readable, and OpenAI has opposed characterizing it as a shift toward completely unintelligible "neural gibberish."
OpenAI's Chief Scientist Jakub Pachocki stated on X that the company has been working to preserve and leverage chain-of-thought monitoring capabilities since its earliest reasoning models, which remains one of the core goals of current research initiatives.
It is worth noting that OpenAI is not the only company exploring this direction. The Information later reported that Anthropic and Google DeepMind are also discussing similar technologies. This suggests that balancing model capabilities with safety and auditability may soon become a broader industry-wide concern in AI.