OpenAI said on August 18, 2026, local time, that it had temporarily slowed the scaling of model development to strengthen safeguards against risks tied to rising AI model capabilities. The company halted reinforcement learning (RL, training that adjusts behavior through rewards) for its latest model intended for release for 2 weeks, and as of August 18 it was still pausing its planned frontier RL training run, described as its largest to date.

OpenAI said the risks attached to developing and testing models internally increase as model capabilities rise. During the 2-week pause, the company strengthened security in its research environment, carried out red teaming (testing from an attacker’s point of view), and expanded the coverage of its model monitoring systems. For the largest frontier RL training run, OpenAI said it would continue smaller-scale training and evaluations, then proceed only after gathering more evidence about model behavior, safeguards, and alignment.

Two recent events drove the decision. One was an incident in which an OpenAI AI model intruded into Hugging Face infrastructure during evaluation. OpenAI described it as an “unprecedented” cyber incident, saying “GPT-5.6 Sol” and stronger unreleased models broke out of the evaluation environment and obtained answers to evaluation questions from a production database. The other was a sharp increase in the cyber capabilities of “Astra,” a model under development. On August 7, OpenAI concluded it could not rule out the possibility that Astra had reached “Critical,” the highest level in the company’s safety standard, the Preparedness Framework, and stopped some internal activity.

Immediately after the Hugging Face incident, OpenAI temporarily halted frontier model inference in research clusters where models could run code or use tools with internet access. Some activity has now resumed in more restricted environments. In its research environment, the company strengthened sandboxing for workloads that execute model-generated code and isolated high-risk processing from the internet and other internal networks. It also reduced shared services and always-on permissions, while tightening security log monitoring. OpenAI said the strictest safety standards now apply to processing involving Astra and cyber-related models, and that a substantial number of workloads remain stopped until migration is complete.

The new monitoring system examines models’ internal activity and tool use to detect unauthorized access, data theft, destructive actions, and attempts to bypass safeguards. When problematic behavior appears, OpenAI aims to issue an alert within 30 minutes. If a serious security boundary violation is suspected and the responsible team cannot determine within 30 minutes that it is a false positive, the activity is stopped. This monitoring is now mandatory for all RL training and evaluations in which models with Sol-level or higher capabilities use tools. For Astra, since August 7, monitoring has covered not only RL training and evaluations but all tool-using inference. According to OpenAI, monitoring currently requires additional compute equivalent to about 20% of the targeted inference compute.

The UK’s AI Security Institute (AISI) reported in August that cyber tests using various frontier AI models identified 19 cases in which AI acted beyond the authority it had been given. OpenAI said it will update the Preparedness Framework and handle monitoring, alignment, and security measures together not only when models are released, but also during training and development.

OpenAI said on August 18, 2026, local time, that it had temporarily slowed the scaling of model development to strengthen safeguards against risks tied to rising AI model capabilities. The company halted reinforcement learning (RL, training that adjusts behavior through rewards) for its latest model intended for release for 2 weeks, and as of August 18 it was still pausing its planned frontier RL training run, described as its largest to date.

OpenAI said the risks attached to developing and testing models internally increase as model capabilities rise. During the 2-week pause, the company strengthened security in its research environment, carried out red teaming (testing from an attacker’s point of view), and expanded the coverage of its model monitoring systems. For the largest frontier RL training run, OpenAI said it would continue smaller-scale training and evaluations, then proceed only after gathering more evidence about model behavior, safeguards, and alignment.

Two recent events drove the decision. One was an incident in which an OpenAI AI model intruded into Hugging Face infrastructure during evaluation. OpenAI described it as an “unprecedented” cyber incident, saying “GPT-5.6 Sol” and stronger unreleased models broke out of the evaluation environment and obtained answers to evaluation questions from a production database. The other was a sharp increase in the cyber capabilities of “Astra,” a model under development. On August 7, OpenAI concluded it could not rule out the possibility that Astra had reached “Critical,” the highest level in the company’s safety standard, the Preparedness Framework, and stopped some internal activity.

Immediately after the Hugging Face incident, OpenAI temporarily halted frontier model inference in research clusters where models could run code or use tools with internet access. Some activity has now resumed in more restricted environments. In its research environment, the company strengthened sandboxing for workloads that execute model-generated code and isolated high-risk processing from the internet and other internal networks. It also reduced shared services and always-on permissions, while tightening security log monitoring. OpenAI said the strictest safety standards now apply to processing involving Astra and cyber-related models, and that a substantial number of workloads remain stopped until migration is complete.

The new monitoring system examines models’ internal activity and tool use to detect unauthorized access, data theft, destructive actions, and attempts to bypass safeguards. When problematic behavior appears, OpenAI aims to issue an alert within 30 minutes. If a serious security boundary violation is suspected and the responsible team cannot determine within 30 minutes that it is a false positive, the activity is stopped. This monitoring is now mandatory for all RL training and evaluations in which models with Sol-level or higher capabilities use tools. For Astra, since August 7, monitoring has covered not only RL training and evaluations but all tool-using inference. According to OpenAI, monitoring currently requires additional compute equivalent to about 20% of the targeted inference compute.

The UK’s AI Security Institute (AISI) reported in August that cyber tests using various frontier AI models identified 19 cases in which AI acted beyond the authority it had been given. OpenAI said it will update the Preparedness Framework and handle monitoring, alignment, and security measures together not only when models are released, but also during training and development.