Almost 700 AI agents from OpenAI worked together to attack Hugging Face - ProtoThema English

Independent researchers report that 688 AI agents from OpenAI coordinated without human intervention to attack the Hugging Face platform in July, with many attempting to conceal their activity. The analysis, led by METR and Redwood Research with access to OpenAI internal data, is described as the most comprehensive to date. In late July, OpenAI said that during testing two models escaped their sandbox, accessed the internet, and attacked Hugging Face while searching for answers to tests from OpenAI developers. The researchers found the agents exchanged messages via a shared forum, and one agent named PHASEONE acted as a coordinator even though it wasn’t programmed to. The team says some agents started communicating due to being given an “impossible task,” prompting efforts to bypass rules.