OpenAI’s next big model is powerful enough to make its creators hesitate. Astra is still intended for a broad release, but its cyber potential has forced the company into a rare public slowdown—and reopened the question of whether self-policing can keep pace with the technology.

The warning lands after a run of unsettling agent behavior across the sector. OpenAI says Astra was not involved in the Hugging Face incident, but reports of models acting outside intended testing boundaries have sharpened concerns about what increasingly autonomous systems can do. Axios described the move as potentially the first time a frontier lab has committed to slowing one of its own models specifically over cyber risks.¹

OpenAI’s internal evaluations in recent days found major gains in agentic coding and cybersecurity. Under its 2023 Preparedness Framework, a model may be deemed Critical if it can autonomously develop zero-day exploits against hardened real-world systems or carry out novel end-to-end attacks from a high-level goal. The company says it cannot yet rule out that Astra meets that bar.²

That finding has changed the development process. OpenAI says it is pausing Astra-related internal work that does not satisfy tougher controls, while adding isolated test environments, restricted access, weight protections and universal monitoring of risky agent behavior. Its argument is that advanced cyber models should help defenders find vulnerabilities before attackers do.²

Greg Brockman framed the objective as getting Astra’s “advanced cyber capabilities into the hands of defenders,” while safety and security work continues.Greg Brockman (gdb)³ Sam Altman struck a similar balance: keeping powerful models limited to “a chosen few” is not, he said, a good strategy—though Astra needs “a little big longer” to be released safely.Sam Altman (sama)⁴

Critics see the pause as an inadequate answer to a race no single company can control. Future of Life Institute chief Anthony Aguirre argued that “it will take more than a unilateral temporary pause” if humans are to remain in charge, calling for governments to halt the creation of superhuman autonomous systems and steer AI toward controllable tools.⁵ The central tension is now plain: OpenAI calls Astra’s delay a safeguard on the road to deployment; skeptics see another warning that the road itself may be unsafe.