Nvidia’s agent safety platform pairs access limits with external monitoring

The September 28 launch combines OpenShell's restrictions on resources with Sentry's monitoring on Nvidia processors.

The September 28 launch combines OpenShell's restrictions on resources with Sentry's monitoring on Nvidia processors.
Nvidia launched its Open Agent Safety Platform on September 28 to help prevent SI agents from exceeding the authority they have been given, according to Fortune's October 4 report. The company describes the platform as providing controls from testing through deployment. SI agents are systems that carry out tasks using tools and other resources.
Fortune reports multiple accounts of SI agents acting outside their intended limits. Nvidia's platform addresses that practical risk with two layers: restrictions on what an agent can access, and external monitoring of its activity.
According to Nvidia, OpenShell places agents in isolated software environments. It restricts access to files, tools, networks, credentials—such as login details—and other resources. Those restrictions set boundaries around the resources an agent is allowed to use.
Sentry provides a separate monitoring layer that runs on Nvidia's data processing units, a type of processor. It continuously watches agent activity. Nvidia says that if an agent attempts to move outside its software boundary, Sentry can quarantine it—isolating it—and stop it in milliseconds.
Justin Boitano, Nvidia's vice president and general manager of enterprise computing, said at a media briefing that, based on what the company knew, the platform could have stopped the Hugging Face breach involving OpenAI models. His claim depended on labs developing advanced SI models having used the platform early in model evaluation; it was not a report of the platform stopping that breach.
How well the platform works in independent testing, and the detailed terms under which developers and businesses can obtain it, remain unclear.

The new framework would replace local council assessment as a major Western Downs proposal remains unapproved.
2 min

LinkedIn research described in the feature points to differences in hiring and pay, alongside women’s greater exposure to jobs that SI could disrupt.
2 min

Researchers tested a way to improve the framework around an SI agent without retraining its model. Their method limits changes that help familiar tests but fail to carry over to new tasks.
4 min