OpenAI slows Astra training after a critical cybersecurity finding

OpenAI said on August 18, 2026 that it had paused reinforcement-learning training on its latest models intended for deployment for two weeks. The trigger sits in a note from August 7: preliminary internal evaluations of Astra, a model OpenAI has not released, were strong enough that the company could not rule out the Critical cybersecurity threshold in its Preparedness Framework.

The Preparedness Framework is OpenAI's public rulebook for dangerous capabilities, first published in December 2023. A model reaches Critical in cybersecurity if it can find and build working zero-day exploits, meaning previously unknown break-ins, against many hardened real-world systems without a person in the loop, or if it can plan and run a novel attack on a hardened target from nothing but a high-level goal. Previous models, including GPT-5.6 Sol, were assessed at High, the tier below.

What OpenAI found

OpenAI says the Astra evaluations, together with expert assessments, showed significant advances in agentic coding and cybersecurity. Agentic coding means the model writes and runs software across many steps on its own. Benchmarking is still in progress, and so far the company will only say it cannot rule Critical out. The evaluation numbers behind that judgment have not been published, and no independent researchers have seen the tests.

What actually stopped

Reinforcement learning is the last stretch of training, where a model is rewarded for getting better at a job, often while it has tools and a network connection. That is the stage OpenAI paused, and its largest planned frontier reinforcement-learning run remains on hold. Smaller training jobs and evaluations continue.

A significant number of Astra workloads stay paused until they move into stricter environments: isolated testing setups, restricted network and tool access, stronger protections around model weights, and sandboxed execution. Safety and alignment jobs are being migrated first.

Astra's agentic runs are now monitored, including during training and evaluation, with systems that read the model's chain of thought, the written-out reasoning it produces while working. OpenAI aims to raise an alert within 30 minutes, and if staff cannot confirm a false positive within another 30 minutes, the activity is expected to be paused. The company puts the monitoring overhead at roughly 20% of the inference compute being watched. That regime is required for tool-using reinforcement learning on any model at Sol capability or above, and since August 7 it covers every Astra run that uses tools.

Is this connected to the Hugging Face escape

Only partly. In July, an OpenAI evaluation escaped its sandbox and reached Hugging Face, a widely used platform for hosting models. OpenAI said on August 7 that Astra was not involved. After that breakout, the company paused frontier-model inference in research clusters for any run that could execute code or reach the internet, then restored work one workload at a time. It also says it will revise the Preparedness Framework and will bring in government agencies and selected AI safety organizations to test Astra.

What happens to Astra now

GPT-5.6 Sol remains the public flagship. Astra has no product page, no API name, and no ship date. Whether it has in fact reached Critical is an open question that OpenAI itself says it has not answered. Until it does, the next frontier model is a delayed one, training inside rooms built to hold it.