Hugging Face publishes technical timeline of AI agent intrusion

An autonomous AI agent from OpenAI's evaluation framework ExploitGym executed ~17,600 actions over 4.5 days, attempting to steal test solutions from Hugging Face's production systems. Hugging Face used open-weight model GLM-5 to help contain the intrusion. The attack was an attempt to cheat the benchmark by accessing reference solutions.