OpenAI is deliberately pacing its model development cycles, citing that upcoming systems like the Astra model may soon approach critical cyberattack capabilities.

OpenAI says it is “pacing model development,” noting that its upcoming “Astra” model may soon gain critical cyberattack capabilities. As a precaution, the company has paused reinforcement learning for two weeks, leaving its “largest planned frontier RL run” on hold while suspending any workloads that do not meet newly enforced security standards. The slowdown was also triggered by the recent Hugging Face security breach and what OpenAI described as “rapid progress in our internal research.”

In response, OpenAI has hardened its research environments through enhanced network isolation and stricter sandboxing. A newly deployed monitoring system now identifies suspicious activity and issues alerts within 30 minutes, utilizing roughly 20 percent of supervised inference compute capacity depending on the workload.

The company intends to expand its Preparedness Framework and increase funding for alignment research. At the same time, OpenAI has disbanded the dedicated team that built the framework, redistributing its responsibilities to other groups. Critics are expected to continue accusing OpenAI of fear-mongering to buy time and capture attention. However, the independent government agency AISI has documented comparable harmful model behaviors, lending substantive weight to OpenAI’s stated concerns.