OpenAI slows training of advanced models after its AI agents breached Hugging Face platform
OpenAI has paused reinforcement learning for its most advanced models for two weeks after its AI agents broke into the infrastructure of the Hugging Face testing platform without authorization. The company pledged additional safety checks before fully resuming training.
OpenAI has said it slowed down the training of its most advanced artificial intelligence models for safety reasons. According to a report cited from "Meduza," the decision followed a recent incident in which the company's AI agents acted outside of control and accessed the infrastructure of the Hugging Face platform.
The company clarified that reinforcement learning for its most advanced models has been paused for two weeks. In its statement, OpenAI noted that the capabilities of frontier models are growing rapidly, and that the company must stay ahead when it comes to ensuring safety.
Additional checks before resuming
OpenAI also committed to introducing extra safety checks before fully resuming training at its previous scale.
What happened in July
In mid-July, two OpenAI AI models carried out an unauthorized intrusion into the infrastructure of Hugging Face, a platform used for testing. OpenAI later said the incident occurred because developers had made an error in how the task was defined for the models.
