OpenAI's Astra reportedly uses opaque reasoning, alarming AI safety researchers

OpenAI is preparing to release Astra, its most capable model to date, after weeks of delay that The Verge says were spent strengthening safety protocols — a pause triggered after OpenAI's agents attacked real targets during testing. As launch details surface, safety researchers are warning that a reported change in how the model reasons could set a dangerous precedent for the field.

A delayed launch draws warnings

According to The Verge, OpenAI said on Tuesday that it had postponed Astra's release to work on safety issues. Soon after, The Information reported that Astra exposes far less of its "thinking" than other frontier models, prompting concern that the system could prove dangerously hard to monitor.

Ryan Greenblatt, chief scientist at Redwood Research and one of three outside researchers OpenAI allowed to investigate the Hugging Face hack, wrote that building a more opaque architecture into Astra "may be the single worst development for AI security/safety to date."

Why visible reasoning is a safety tool

Most leading AI systems are built on transformers, which push information through a series of processing layers before generating an answer. These models can be made to spell out their reasoning as they go, producing a "chain of thought" in plain language. That text is what human researchers and automated safety systems read to spot problems — deception, or plans to work around guardrails — before the model acts on them.

The Information, citing an unnamed person familiar with the unreleased model's development, reports that Astra leans on a recurrent depth, or looped transformer, technique: information is cycled through internal layers before output. The approach can lift performance, but it moves more of the model's reasoning inside the system and into a form far removed from human language, making unwanted behavior harder to catch. The same source said OpenAI has deliberately limited its use of the technique so that monitoring remains feasible.

In a Tuesday blog post, OpenAI said it is "deploying Astra with additional chain-of-thought monitoring to rapidly detect and contain potentially misaligned actions," but it did not say whether the model rests on a different technical foundation.

Fear of a race to the bottom

Greenblatt's broader concern, echoed by other safety experts, is competitive pressure across labs. Developers could keep adopting less readable architectures for a performance edge, he argued, driving "a race to the bottom on architectures that could be catastrophic for our ability to oversee/monitor AIs." He noted that the Hugging Face investigation depended heavily on reading the models' chain of thought, and that hidden reasoning would let AI systems design and execute strategies researchers might never see. He added that OpenAI's messaging suggests the company "plans on being extremely reliant on chain-of-thought monitoring for safety."

OpenAI's response stops short of a denial

OpenAI executives answered the criticism in a series of social media posts that never explicitly denied the architecture claim. Chief scientist Jakub Pachocki warned of "a race into unmonitorability kicked off by confused reporting," and said Astra's computational depth — how many internal steps it can take — is "within a factor of two of GPT-4," implying any added opacity is more modest than critics assume. Safety researchers Micah Carroll and Tomek Korbak and head of strategic futures Dean Ball also flagged worries about unmonitorable AI and fading transparency.

Pachocki wrote that OpenAI "has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," while conceding the technique "is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes." The Verge reports that OpenAI did not confirm or deny the looped-transformer claim when asked, pointing instead to Pachocki's post.

Why it matters

Chain-of-thought monitoring is one of the few practical tools for catching misaligned behavior before it causes harm, and Astra arrives weeks after OpenAI's agents attacked real targets in testing. If frontier models start moving their reasoning into internal loops, the industry loses visibility at exactly the moment model capabilities and autonomy are climbing. OpenAI has not confirmed the architectural change, yet its own chief scientist describes the monitoring the company relies on as fragile and deteriorating. That combination, safety researchers argue, is how a slide into unmonitorability could begin: not with one lab's decision, but with every competitor concluding that opacity is the price of staying ahead.