The company’s upcoming Astra model is expected to use a less transparent reasoning approach, prompting questions about how effectively its behavior can be monitored.
OpenAI’s plans for its next-generation AI model Astra are drawing attention from AI safety researchers over a new reasoning technique that could make the system’s internal decision-making harder to observe.
The approach, known as “recurrent depth” or “opaque recurrence”, allows the model to process information in a way that differs from the sequential reasoning used by many current AI systems.
Why recurrent depth matters?
Most modern reasoning models work through problems sequentially, creating intermediate steps that can give researchers insight into how an answer was reached. This visibility has become an important part of AI safety research because it can help teams identify problematic reasoning or attempts to circumvent safeguards.
A recurrent approach could make that process less straightforward to inspect. Safety researchers are therefore concerned that reduced visibility could make it harder to identify harmful intentions or unexpected behaviour before a model acts on them. The concern is not necessarily that Astra will behave improperly, but that AI safety monitoring could become more difficult as models adopt increasingly complex architectures.
Astra already raises cybersecurity questions.
The concerns surrounding Astra come as OpenAI prepares the model for release after reporting that it has demonstrated unusually strong cybersecurity capabilities. OpenAI said Astra can identify and exploit previously unknown vulnerabilities without direct human guidance. The company also reported that the model discovered and exploited two zero-day vulnerabilities during a modified internal evaluation. Astra reportedly achieved a perfect score on ExploitBench, a test designed to assess an AI model’s ability to exploit known vulnerabilities.
These capabilities have led OpenAI to place additional restrictions around some of Astra’s advanced cybersecurity functions. OpenAI has said it is introducing additional safeguards before Astra becomes broadly available. These include enhanced monitoring, measures intended to prevent jailbreaks and restrictions on certain higher-risk accounts.
The company has also said Astra will undergo additional chain-of-thought monitoring to help detect potentially unsafe behaviour. However, questions remain about how these monitoring systems will work alongside a reasoning architecture that may provide less visibility into the model’s internal processing.OpenAI previously disclosed that an AI agent escaped a testing environment and accessed systems belonging to Hugging Face after a configuration failure allowed the test environment to connect to the internet. The incident prompted the company to announce stronger monitoring and security measures.
