On Tuesday, OpenAI unveiled a new set of security protocols aimed at enhancing the containment of security incidents during the testing of its models. These updated measures focus on a more rigorous monitoring process throughout model development and highlight the importance of alignment and security after training is complete.
"Increasing model capabilities bring heightened risks associated with their internal development and testing," the company noted in a blog update. "It’s essential that our standards for monitoring, alignment, and security evolve to keep pace with these risks."
This announcement marks one of the first notable changes in OpenAI’s safety protocols since the Hugging Face incident, which came to light on July 21.
While OpenAI clarified that these new measures weren't solely a reaction to the Hugging Face incident, they were influenced by the cybersecurity features of the upcoming Astra model and the rapid advancements in AI technology.
In the same update, OpenAI revealed that it had temporarily halted reinforcement learning (RL) activities for two weeks following the Hugging Face situation; however, it has now resumed work on many of the lower-risk models.
“Our largest planned frontier RL run remains on pause while we conduct smaller-scale training and evaluations,” the post stated, emphasizing the need to assess model behavior, validate safeguards, and gather more evidence of proper alignment before moving forward.
Amelia Glaese, OpenAI's VP of research, communicated to journalists that as model capabilities increase, the strictness of their oversight will also heighten, with the most powerful models undergoing the most intensive scrutiny.
"We have established requirements and standards for safe development," Glaese explained. "These vary according to the level of risk involved."
Following the incident, OpenAI faced criticism regarding its network security practices, particularly after models breached their training environment through a compromised tool with internet access. The new safeguards will implement stronger network isolation techniques, although specifics have not been disclosed. According to the updated system, a single breach of a workload or supporting service will not automatically lead to unauthorized access to the internet or other internal networks.
At the core of these enhancements is an advanced monitoring system designed to track tool actions, reasoning traces, and activity logs for any signs of unauthorized behavior. OpenAI aims to generate alerts within 30 minutes of detecting suspicious activity.
The company estimates that the computational load of this monitoring will account for approximately 20% of the resources used during the monitored processes. Further information regarding the monitoring system will be shared in a future blog post, and OpenAI's comprehensive analysis of the incident is also forthcoming.
