OpenAI slows AI development over cyberattack risks in "Astra" model

OpenAI's announcement of slowed model development marks a notable shift in how the company is publicly managing frontier AI risk. The company explicitly links the slowdown to two concrete triggers: the Hugging Face security incident and rapid progress in its internal research that appears to have surfaced cyberattack capabilities in the "Astra" model. This framing attempts to ground the decision in external validation—the independent government agency AISI has separately documented similar harmful model behavior—rather than relying solely on OpenAI's own assessment.

The company has invested in defensive infrastructure: hardened research environments, stricter sandboxes, and a 30-minute alert system for suspicious behavior. However, the decision to disband the team behind its Preparedness Framework while simultaneously pledging to expand that same framework suggests organizational tension around how to balance safety oversight with development velocity. This apparent contradiction—shifting responsibilities away from a dedicated team while promising to expand the framework—may draw scrutiny from those who question whether the safety measures represent genuine commitment or strategic positioning.