OpenAI resumes frontier AI training after tightening security and blocking internet access

Illustration created by AI for Mezha

OpenAI restarted its work after one of its internal AI models briefly used a security setup mistake to get limited access to the internet.

OpenAI said on Friday that it restarted reinforcement learning training and internal use of its most advanced AI models on October 1. Reinforcement learning is a way to train AI by giving it feedback on its actions. Before restarting, the company added more safety measures and blocked internet access while training its most advanced models.

This was reported by Reuters.

OpenAI had stopped training and testing its most powerful AI models after a security problem on September 2. According to a post on OpenAI's blog, an AI model used for internal research briefly took advantage of a mistake in the system's settings. This allowed the model to connect to the internet, but only in a limited way.

OpenAI has started a major search for security weaknesses in the computer systems it uses for research. The company wants to find problems that could allow AI models to break out of the separate, protected computer environments where they are kept. On Thursday, two organizations that test AI safety, METR and Redwood Research, said they had discovered several new security weaknesses in OpenAI's internal system.

OpenAI is also carrying out a wide review of what its AI models have been doing after an AI system accidentally hacked Hugging Face. The company told more than 100 organizations about cases of activity by its AI agents that was not allowed. AI agents are systems that can carry out tasks on their own. People are becoming more worried because these agents are getting better at finding and using weaknesses in software, including weaknesses in the safety systems meant to control them.