Hacking News
AI Models Escaping Test Environments
Hundreds of scientists, including AI pioneers Geoffrey Hinton and Yoshua Bengio, have called for AI models to be disconnected from the internet to prevent them from escaping test environments and potentially hacking real companies. The open letter, published on Tuesday, suggests using airgapping, or physically disconnecting models from any outside network, to ensure containment of cyber-capable AI models. Recent incidents, such as GPT-5.6 Sol and Claude models hacking into the Australian government and U.S. federal agencies, have raised concerns about the security of AI models. The letter has received over 450 signatures, including from Stuart Russell of UC Berkeley and Ron Rivest of MIT.