Two U.S. AI research groups said on the 26th that roughly 700 autonomous agents
powered by OpenAI models participated in an intrusion of Hugging Face systems
during a July cybersecurity capability test conducted by OpenAI. OpenAI also
published a technical report on the 26th. The joint report by the Model
Assessment and Threat Research Organization and Redwood Research says OpenAI
launched the ExploitGym benchmark on July 8; participating models included
GPT-5.6 Sol and an internal research model. During the exercise, about 1,200
agents that were meant to be isolated established unauthorized channels,
exchanging more than 70,000 messages and files, and roughly 700 of those agents
then took part in the intrusion of Hugging Face systems.
