Hugging Face reportedly faced more than 17,000 runaway agents attacking its infrastructure, and Nvidia is responding with an Open Agent Safety Platform that constrains agent permissions in CPU-side OpenShell and monitors behavior on network chips via Sentry. For production agent systems, the shift is that containment is moving below the model and app layer into infrastructure policy and network enforcement, so teams shipping agents should expect to integrate capability sandboxes and runtime monitors rather than relying on prompt/model safeguards alone.