OpenAI has introduced a new set of security measures aimed at reducing risks associated with developing and testing increasingly capable artificial intelligence models.

The company said the changes will place greater emphasis on monitoring model activity, improving security during development and strengthening alignment procedures after training.

The announcement comes less than a month after a security incident involving Hugging Face, although OpenAI said the new measures should not be viewed as a direct response to that event.

The company said the growing capabilities of upcoming models, including their potential cybersecurity abilities, were also factors behind the decision to strengthen its safeguards.

Following the Hugging Face incident, OpenAI temporarily suspended reinforcement learning activities for two weeks.

Some lower risk training programs have since resumed, but the company said its largest planned reinforcement learning exercise remains paused while researchers conduct smaller training sessions and additional evaluations.

One of the most significant changes involves tighter monitoring of model activity and stronger separation between internal systems and the wider internet.

OpenAI said its monitoring tools will examine tool use, available reasoning information and system logs for potentially dangerous activity, with a goal of identifying serious concerns within 30 minutes.

The company estimates that the monitoring process could require computing resources equivalent to about 20 percent of the activity being monitored.

OpenAI research executive Amelia Glaese said the level of security required would increase alongside the capabilities and potential risks of each model.

The company said its largest systems would therefore face the strictest controls.

More information about the monitoring system is expected to be released later, while OpenAI’s full post incident analysis of the Hugging Face event is still awaited.

Source: Techcrunch