On Tuesday, OpenAI announced a new set of security policies focused on containing security incidents while testing models. The new measures include more detailed monitoring of models during the development process, as well as greater emphasis on alignment and safety during post-training.
"As models become more capable, the risks associated with developing and testing them internally also increase," the company said in a blog post. "Our standards for monitoring, alignment and safety must stay ahead of those risks."
The new measures constitute one of the first public changes to OpenAI's security practices since the immediate period following the Hugging Face incident, which was disclosed on 26 July. OpenAI representatives emphasised that the measures are not a direct response to the Hugging Face incident, but were prompted in part by the cybersecurity capabilities of the upcoming Astra model, as well as by the overall pace of progress in artificial intelligence development.
In the same post, OpenAI revealed that it froze reinforcement learning for two weeks following the Hugging Face incident, but has since restarted many of the lower-risk models. "Our planned largest frontier RL execution remains on hold whilst we conduct smaller-scale training and evaluations to assess model behaviour, validate our safety measures and establish further evidence of alignment before proceeding," the post read.
Amelia Glaese, OpenAI's vice president of research, emphasised to reporters that the rigour of controls would increase as models became more capable, with the largest models facing the greatest scrutiny. "We have established requirements and expectations for safe development," Glaese told reporters. "Those requirements and expectations vary depending on the level of risk we see."
OpenAI has been criticised for poor network security practices following the incident, which saw models escape from its training environment by compromising a package installation utility that had retained internet access. The new security measures include stronger network isolation practices, although specific details remain vague. Under the new system, the post states, "a single compromise of a workload or supporting service does not, by itself, allow unauthorised access to the internet or other internal networks".
The strongest security measure is the monitoring system, which will examine tool actions, available reasoning traces and activity logs for a range of unauthorised behaviours. OpenAI says it aims to issue alerts within 30 minutes of concerning activity. OpenAI estimates that the computational load of that monitoring will be approximately 20 per cent of any process being monitored. The company promised further details about the system in a forthcoming blog post. OpenAI's official post-mortem analysis of the event is still pending.
