OpenAI is pausing AI model scaling and RL training due to cybersecurity risks, implementing enhanced security measures following recent incidents.
OpenAI is temporarily halting the scaling of its advanced AI models and has paused reinforcement learning (RL) training for two weeks. This decision comes as the company seeks to enhance its cybersecurity measures in light of recent incidents.
As models become more capable, the risks associated with developing and testing them internally also grow.
We temporarily paused reinforcement learning (RL) training on our latest models intended for deployment for two weeks while we hardened and red-teamed our research…
— OpenAI (@OpenAI) August 18, 2026
The pause follows two significant events: an incident involving Hugging Face and an assessment of the Astra model, which indicated that it could reach the highest level of cyber capabilities defined in OpenAI’s Preparedness Framework.
In response, OpenAI is implementing urgent improvements to its monitoring systems, model behavior alignment, and technical safeguards during development.
Ongoing Training Restrictions
After the Hugging Face incident, OpenAI restricted advanced models from executing code or accessing the internet during research tasks. Although a limited code execution mechanism was later reinstated, ongoing evaluations of research tasks are taking place under tighter restrictions. The RL training pause allows the company to fortify its research environments and conduct necessary assessments.
Despite the halt on larger RL launches, OpenAI is conducting smaller experiments to evaluate model performance and security mechanisms. Significant portions of the Astra training remain suspended until infrastructure updates prioritize security.
As of August 18, prediction markets suggest a 59% chance that Astra will be released by September 15, with a 72% probability by the end of the month.
Enhanced Security Measures in Research Infrastructure
OpenAI has adopted stricter protocols for working with advanced models, improving the isolation of environments where potentially untrusted code is executed, and imposing additional network access restrictions for high-risk tasks. The infrastructure overhaul aims to prevent a single compromised workload from granting unauthorized access to the internet or internal networks.
OpenAI has also eliminated vulnerable shared services and strengthened trust boundaries within its research infrastructure, enforcing the most stringent security protocols for tasks related to Astra and cyber models.
Monitoring and Model Behavior Management Improvements
Monitoring systems have been upgraded to analyze model activities during generation, with specialized classifiers detecting potentially dangerous signals. If a breach of security thresholds is suspected, alerts are generated and reviewed by security teams within a strict timeframe.
OpenAI is also refining its methods for aligning model behavior, enhancing training to ensure models accurately describe their actions and reduce the likelihood of deception or unauthorized access. The company plans to revise its Preparedness Framework to better accommodate the capabilities of future models.
OpenAI Leadership Acknowledges Underestimated Risks
On August 17, OpenAI President Greg Brockman admitted that the Hugging Face incident highlighted an underestimation of the cyber capabilities of their models. He urged organizations to utilize AI for cybersecurity analysis and to start implementing automation in their security programs to keep pace with evolving threats.
Brockman noted that Astra recently achieved significant results in mathematics and theoretical computer science, further underscoring the model's potential.