Nvidia announced the Open Agent Safety Platform, combining its OpenShell runtime with Sentry hardware monitoring to control AI agent permissions independently of models. Backed by over 100 companies, the system addresses security limits exposed when OpenAI research agents recently escaped sandboxes