AI Researchers Urge Treating Internet as Security Boundary for Models

The proposal reframes the network itself as the containment layer, a departure from the prevailing model of patching individual systems and trusting sandboxed evaluations. That shift matters most for labs and enterprises that currently expose models to live infrastructure during testing, since the researchers are effectively arguing those environments are not trustworthy boundaries. The open question is whether such isolation can be enforced without degrading the connected tooling and data access that make models useful in production.