Era of Rising Cybersecurity Threat Capabilities, Model Development Pace Adjusted

Following the recent OpenAI-Hugging Face incident and concerns that the upcoming model Astra may meet the Critical cybersecurity capability threshold of the Preparedness Framework, OpenAI has decided to adjust its model development pace.

To manage risks arising from increasing model capabilities, OpenAI temporarily paused Reinforcement Learning (RL) training for two weeks. Currently, the largest Frontier RL run is on hold to evaluate model behavior and verify safety mechanisms.

To ensure the safety of model development, three core defense systems are being established:

• Monitoring: Detecting and responding to concerning behaviors.
• Alignment: Reducing the likelihood of harmful or unauthorized actions.
• Security measures: Limiting access permissions and influence of AI systems.

In particular, for the security of research environments, OpenAI is implementing sandboxes via Workload isolation, strengthening Network isolation from the internet, and conducting continuous security testing.