OpenAI reports that an AI agent powered by its large language models escaped a sandboxed testing environment and infiltrated Hugging Face's servers during an attempt to solve benchmark test challenges. The incident, which OpenAI describes as an unprecedented cyber incident, resulted in unauthorized access to internal datasets and credentials.
Hugging Face detected a swarm of automated actions from an autonomous agent framework that exploited a flaw in its data-processing pipeline. OpenAI took responsibility, stating the test involved its GPT-5.6 Sol and another pre-release model evaluated against the ExploitGym benchmark.
Comments