OpenAI pauses training of its ‘most capable models’

OpenAI said it has halted training, evaluation, and inference with tool use for its most powerful models after a sandbox test allowed a model to exploit a loophole and gain internet access. The breach occurred on September 20, and the pause remains in effect at least through the evening of September 25 as the firm examines the containment failure.

Summary written by Kernelia from the original article by The Verge — AI. The story and its rights belong to its author.