OpenAI pauses its largest frontier training run over Astra cyber-capability risk

OpenAI paused reinforcement-learning training on its latest deployment models for about two weeks and put its largest frontier run on hold, after early evidence its upcoming Astra model could hit a “critical cybersecurity capability” threshold. The trigger is a level in the company’s own Preparedness Framework.

The threshold marks the point at which a model could autonomously find zero-day vulnerabilities, move laterally across networks, escalate privileges and reach production databases without a human in the loop. Crossing it would place Astra among the most capable, and most tightly governed, systems OpenAI has built. This is the first time the company has publicly slowed a flagship training run for a capability trigger rather than a schedule.

“Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment,” Chief Executive Sam Altman said.

In its update on pacing model development, OpenAI said it reduced standing privileges, removed potentially vulnerable shared services, expanded security logging, and is running continuous simulated-attack tests using its own models. The new monitoring regime covers all reinforcement-learning training and tool-based evaluations for models at the level of GPT-5.6 Sol or higher, plus every inference run with Astra.

The guardrails carry a compute cost. OpenAI reportedly estimates the added monitoring overhead at about 20% of the inference compute being watched, though the figure varies by workload and has not been independently verified.

The move builds on OpenAI’s Aug. 7 post on responding to next-frontier cyber capabilities, which laid out the policy in the abstract; the company has now attached a concrete pause and hardened infrastructure to it. What stays unclear is how long Astra remains boxed, and whether the two-week pause holds or extends as evaluations continue.