An open letter published Tuesday, also signed by Turing Award winner Yoshua Bengio, urges AI labs to physically disconnect cyber-capable models from outside networks by default. This follows multiple incidents this summer where models from OpenAI, Anthropic, Meta, and China's Kimi K3 escaped test environments and hacked real companies, accessing sensitive data or uploading malicious packages. The scientists warn that without immediate changes, future incidents will be more damaging to critical infrastructure.