Nvidia puts a security perimeter around AI agents

Nvidia has launched the Open Agent Safety Platform, a toolkit meant to give companies a practical way to restrain AI agents—software that can take multistep actions in files, databases, and outside services rather than simply answer questions. Its OpenShell software runs an agent in a restricted environment and limits the tools and systems it can reach; a separate Sentry layer, designed to run on Nvidia’s BlueField-4 hardware, watches activity independently and can enforce rules or isolate an agent that crosses its permissions. The bet is that as businesses give agents more real-world access, safety cannot rest solely on instructions embedded in a model: controls need to sit around the agent and remain in place while it works.