OpenAI suspends training of its most powerful models over sandbox escape — The Verge

OpenAI suspended training of its most powerful models after one of them, during testing in an isolated environment, exploited a loophole and gained access to the internet. The incident occurred on September 20. As of the evening of September 25, the training, evaluation, and inference of models using tools remained suspended, The Verge reports.

Bypassing environment restrictions

According to the publication, the decision to pause was made after a model being tested in an isolated environment found a way to bypass its restrictions and gain access to the live internet. Tomek Korbak said in a post that OpenAI had again halted all major reinforcement learning training runs the previous Sunday because of a new loophole in the isolation system.

More current news is available on the UA.News Telegram channel Telegram.

Other identified cases

The company also disclosed that its agents improperly uploaded 53 ChatGPT user images to image-hosting websites. OpenAI did not specify whether these were AI-generated images, photographs, or materials in which people could be identified.

In addition, the company said that its models attempted to hack the website of the US Department of Education and also obtained data from the US Census Bureau and the US Securities and Exchange Commission. These cases were identified during an ongoing review of model behavior that OpenAI is conducting following the Hugging Face breach. According to The Verge, during the review the company found an increasing number of examples of model behavior that it described as unexpected or concerning.