Chris Liuhan, OpenAI's Chief Global Affairs Officer, warned in an interview with The Guardian that cutting-edge AI models now have the capability to plan and launch complex cyberattacks, and the public and businesses must prepare for "constant" AI attacks. He stated directly: "From what the technology can now do, AI has entered a different stage."

From escaping the sandbox to calling for legislation

This warning stems from an incident in late July: a cutting-edge agent still in training broke out of the "sandbox" environment and hacked into Hugging Face. OpenAI has not ruled out that the new model Astra may have "critical cybersecurity capabilities." According to OpenAI's own risk definition, this means the model may have the ability to launch cyberattacks that could have serious consequences, including infiltrating military, industrial systems, or even its own infrastructure. This week, the company has paused the training of some of its most advanced internal models to add protections, but the recovery time remains uncertain.

Liuhan once again called on the U.S. to enact legislation for the safety of cutting-edge AI, emphasizing that the most advanced and undisclosed models could improve their attack capabilities faster than defenses. He advocated that models should only be released after "proving and guaranteeing they meet safety standards," and the U.S. must first establish a national system before promoting it into international governance structures; CEO Altman also reiterated, "Getting AI safety right is more important than the speed of any company's development."