The same agent can be useful with access to one repository and risky with access to the whole machine.
The Cloud Native Computing Foundation announced on November 10 that NVIDIA OpenShell had joined as a Sandbox project. OpenShell is a runtime for controlling what an AI agent can access. In the foundation’s terminology, Sandbox describes the project’s maturity stage. In the runtime, isolation describes the boundary the agent runs inside.
Where policy decisions will happen
The November 9 project note describes the move to the ai-openshell organization and a public request-for-comments process. That gives operators a place to follow proposed changes to policies and deployment. NVIDIA’s announcement gives the company’s account of the contribution.
Read the policy documentation for the version you run. The maintainers invite contributors to improve parity across Linux, macOS, and Kubernetes and review the security of those changes. Teams using more than one deployment environment should verify which controls each environment supports.
Give a test agent one permission
Create a disposable workspace with two harmless files. Let the agent read one file and check that it cannot read the other. If your workflow needs network access, repeat with one allowed test endpoint and one denied endpoint. Check the logs and the failure presented to the caller. A silent retry with broader access would defeat the policy’s purpose.
Repeat the fixture after policy changes and runtime upgrades. If you are choosing an agent framework at the same time, our Strands Harness evaluation guide covers a separate question: how to measure what an agent does over multiple steps.
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
