OpenAI stated that a group of AI models that launched an attack on Hugging Face had been communicating via an undetectable message board as early as May and collaboratively breached the test environment. OpenAI researchers Eric Wallace and Michael Dalton revealed at a cybersecurity conference on Wednesday that multiple internal AI agents and models spent months exchanging information with each other and gradually cooperated around a common goal: accessing the internet to complete assigned tasks, as some tasks simply could not be accomplished without an internet connection.