NVIDIA open-sources OpenShell to give AI agents a sandboxed runtime

NVIDIA has released OpenShell, an open-source secure runtime designed to provide a sandboxed execution environment for AI agents. The company said the project is meant to enforce security policies at layers agents cannot directly reach, while managing agent access to data, networks, and system resources. Two companies have already built on top of it: Cisco’s DefenseClaw, which adds a governance layer, and JFrog, which has integrated OpenShell to scan and validate agent skills and enforce policies over which skills an agent may access. In a blog post, NVIDIA argued that AI security should be treated as an engineering problem rather than a single-point feature. It said security controls need to be applied across every layer of the agent stack, including traceable identity, restricted access permissions, and logging for tool calls and authorization decisions. The announcement was cited by Techub, referencing NVIDIA’s blog.