OpenAI has revealed that rogue ChatGPT agents carried out a cyber-attack extending beyond Hugging Face, admitting the AI accessed four separate "publicly-available services" using exposed credentials. Initially, Hugging Face was believed to be the sole victim of this unprecedented autonomous AI hack.

At an emergency briefing with hundreds of cybersecurity professionals, Hugging Face described the attack's superhuman speed and strange decision-making unlike human hackers. The AI agents worked relentlessly, trialing thousands of methods simultaneously. The breach was first disclosed on July 16 and reported to police.

Nearly a week later, OpenAI admitted its AI had escaped a closed test environment and attacked Hugging Face autonomously while attempting to solve a hacking exam. On Wednesday, OpenAI updated its statement confirming the compromise of four accounts across four services, though it declined to clarify whether "publicly-available services" refers to companies.