Agent security needs hard limits outside the model

AI agents need enforceable boundaries that still hold when the model makes a bad decision, Nvidia says in a new security blueprint. Instructions can guide an agent, but the environment around it must independently control which files, networks, processes and services the software can reach.

The approach assigns every agent a traceable identity with credentials limited to its task. Updating a customer record, for example, should not automatically grant permission to export that record. An agent may request broader access, but should not be able to approve its own request. Human authorization remains necessary for consequential actions and permission changes.

Nvidia points to OpenShell, its open-source runtime for sandboxing agents and enforcing access policies outside their control. Partners in the Open Secure AI Alliance are extending that layer: Cisco’s DefenseClaw adds governance, while JFrog scans agent skills and restricts which ones can run. The company also recommends protected logs of tool calls, policy decisions and outcomes, plus repeated attack tests after material changes.

This is Nvidia’s engineering position, not proof that the listed tools make an agent safe. Its useful standard is evidence: teams should demonstrate that controls block credential theft, unauthorized transfers and monitoring interference, assign an owner to each deployment, and turn every discovered failure into a repeatable regression test.