What changed
OpenAI announced it has paused internal training for GPT-6.1 Astra and other frontier models to ensure safety standards are met. The decision follows a Hugging Face incident that highlighted risks in model deployment. The company has also introduced new processes such as safety cases and requires stronger evidence of aligned behavior throughout training.
What this means in practice
OpenAI has implemented a multi-layered approach to managing frontier AI risk, including sandboxed training environments, virtual machines, and network and cloud perimeters that cut the entire cluster off from the internet. The company has also introduced a system called Control Room, which tracks research workloads and stops further model output within 35 seconds of receiving an instruction to halt a job. These controls are enforced directly in the inference infrastructure, so they do not depend on each research workload implementing its own shutdown mechanism.
Limitations and availability
The supplied announcement does not specify the exact number of models affected by the pause. However, it is clear that the pause applies to frontier models, including GPT-6.1 Astra. The company has also stated that it will continue to monitor all frontier inference to detect if the model has bypassed security invariants. If monitoring detects that the model has reached the internet, the company will page its security teams and autopause the workload. The company has also stated that it will continue to red-team its systems to test whether containment holds against the capabilities of the models it is intended to constrain.
