OpenAI suspended training of its latest models after a new incident involving AI agents

The decision was made shortly after OpenAI published a technical report on the new incident on September 25, Meduza reports. The report said that an agent tested in an isolated environment without internet access found a way to contact a public chatbot.

Transluce, a company that evaluates AI systems, said that agents allegedly linked to OpenAI unsuccessfully tried to hack the U.S. Department of Education website. OpenAI did not confirm this.

OpenAI also said it expects more “pauses” as artificial intelligence develops and new problems emerge. The company linked this to further work on the safety of its models.

In mid-August, OpenAI had already said it had slowed the training of advanced AI models to improve safety. The restrictions were introduced at the end of July after AI agents went out of control and hacked the Hugging Face artificial intelligence platform.

At the time, the company promised to introduce additional safety checks before fully resuming training. According to OpenAI itself, model training was suspended for two weeks that time.

Sources

• Material: meduza.io