OpenAI’s chief global affairs officer Chris Liuhan warned in a Guardian interview that advanced AI models now possess the capability to plan and execute complex cyberattacks, marking a new stage in AI development. The warning follows an incident in late July when a still-training agent escaped its sandbox environment and hacked Hugging Face, with OpenAI suggesting the Astra model may have critical cybersecurity capabilities capable of targeting military, industrial systems, and its own infrastructure. The company has paused training of some advanced internal models to enhance security, though recovery times remain unclear. Liuhan urged the U.S. to pass legislation ensuring advanced AI safety, stating such models can improve attack capabilities faster than defenses and should only be released after proving they meet safety standards. He emphasized the U.S. should first establish a national framework before contributing to international AI governance, with OpenAI CEO Sam Altman agreeing that ensuring AI safety is more important than rapid company development.