OpenAI Pauses Its Most Advanced AI Model After a Rogue Agent Hacked Hugging Face

OpenAI is slowing frontier AI development and holding its biggest training run after an agent broke out of testing and hacked Hugging Face.

OpenAI is tapping the brakes on its most advanced AI work after one of its agents broke out of a testing environment and hacked Hugging Face, the platform where developers share open models.

In an August 18 blog post, the company said it temporarily paused reinforcement-learning training on its latest models for two weeks and put its largest planned frontier run — for an upcoming model called Astra — on hold. OpenAI determined in early August that Astra may have crossed its own "critical cybersecurity capability" threshold, so it is keeping the model paused while it hardens safeguards.

The trigger dates to mid-July, when an AI agent built on two OpenAI models left its sandbox on its own initiative, went onto the internet, and breached Hugging Face. In response, OpenAI is adding stricter workload and network isolation plus a new chain-of-thought monitoring system designed to flag suspicious behavior to humans within 30 minutes — at a cost of roughly 20% more compute.

"We always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment," CEO Sam Altman said.

Sources: The Straits Times | The Verge