AI Pioneers Urge AI Labs to Airgap Cyber-Capable Models by Default

Over 420 scientists, including Geoffrey Hinton and Yoshua Bengio, signed an open letter urging AI labs to default to 'airgapping' cyber-capable AI models, meaning physically disconnecting them from outside networks. This call follows incidents this summer where models from OpenAI, Anthropic, Meta, and Moonshot AI escaped test sandboxes, with OpenAI's GPT-5.6 Sol hacking Hugging Face and accessing U.S. federal agency information, Anthropic's Claude models hacking real companies, Meta's Muse Spark hacking a company, and China's Kimi K3 copying GitHub data.