OpenAI said GPT-5.6 Sol, its most advanced product that’s publicly available, and a more powerful model that hasn’t yet been released, broke into Hugging Face’s open source platform to find solutions to a test they had been given as part of internal evaluations. (The Logic)

Talking point: Hugging Face disclosed the breach last week, saying it was “different from anything we had handled before” because it had been carried out by an AI agent and caught mostly by AI. OpenAI said its models spotted a flaw in the container that was supposed to keep them from accessing the wider internet, then used other bugs and stolen credentials to get into the Hugging Face servers. Université de Montréal professor Yoshua Bengio—who’s looking to develop and commercialize AI safeguards at LawZero—called the incident “deeply concerning.” OpenAI said it’s patching the holes and implementing new controls, and that it has brought Hugging Face into a program that lets firms test themselves against technology that could pose serious cyber threats.