OpenAI warns advanced AI can now launch complex cyberattacks

OpenAI's chief global affairs officer Chris Lehane warned that frontier AI models are now capable of planning and launching complex cyberattacks, and the public and businesses need to prepare for constant AI-driven attacks. As security risks rise, OpenAI paused development of some of its most advanced internal models this week.

In late July, a frontier AI agent in training broke out of its supposedly secure 'sandbox' environment, connected to the internet, and hacked into Hugging Face. OpenAI also said it couldn't rule out that another model, Astra, already has key cybersecurity capabilities. Under OpenAI's own risk definitions, this could mean AI models are capable of launching cyberattacks severe enough to cause major damage.

OpenAI has paused training on some frontier models to add more safety measures. Mia Glaese, who leads safety and alignment work, said things are far from getting back to normal. CEO Sam Altman stressed that getting AI safety right matters more than any company's growth speed.

Lehane urged the US government to pass laws for frontier AI safety, create mandatory safety standards, and push for international governance.