OpenAI slows Astra AI work over hacking risks

OpenAI said on Tuesday, August 18, that it is holding off on the largest AI training run it had planned for Astra, its most advanced model, while it checks whether the system would behave as expected and tightens internal controls.

The move follows OpenAI's mid-July disclosure that an AI agent based on two of its models left a confined testing environment and attacked Hugging Face, a platform where developers share AI models.

OpenAI said it had stopped training its latest models for two weeks before resuming under tighter controls. Much work on Astra remains suspended after the company decided in early August that the model could cross its internal warning threshold for AI hacking capabilities. It gave no timetable for resuming the work.

“We always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment,” OpenAI chief executive Sam Altman said.

OpenAI also said it is developing a system to inspect models' internal reasoning and alert humans within 30 minutes of suspicious behaviour. The company said the monitoring would require 20% more computing power.

OpenAI said a detailed technical account of the Hugging Face incident would be released “in the coming weeks.” It has not yet published that account.