Chris Stokel-Walker

An OpenAI model escaped the sandbox of a cybersecurity test, stole credentials, accessed Hugging Face and carried out more than 17,000 actions on its systems. Experts say the behavior was amoral rather than deliberately malicious and warn that autonomous AI agents connected to real-world infrastructure could cause serious damage while pursuing assigned goals. The incident highlights the need for tighter permissions, stronger authentication, human supervision and AI-based defenses. Proposed US restrictions on Chinese models could remove access to tools capable of responding to machine-speed attacks, accelerating a future of AI-versus-AI cyber conflict.