On 28 September, Nvidia launched its Open Agent Safety Platform. It is intended for AI agents: programs that can plan several steps and use tools to carry out work.
The platform includes OpenShell, open-source software that runs an agent in a restricted environment. The operator decides which files, networks and services the agent may reach. Passwords and keys stay outside the agent’s workspace. Sentry, another part, watches behaviour from separate Nvidia hardware and stops attempts to cross those boundaries.
OpenShell is available now under the Apache licence. It requires technical setup: Linux or an Apple-silicon Mac, plus Docker or a similar tool. Sentry is a reference design for larger systems and requires specialised hardware. This is not a security app for a home computer.
Nvidia says Sentry can isolate an agent within milliseconds. That is the company’s own claim and does not mean every risk has been solved.
The basic idea is simple and useful: an AI that can take action should begin with the least access possible. New permissions should be visible and require deliberate approval.