NVIDIA OpenShell Framework Addresses AI Security Across Agent Stack

NVIDIA has introduced the OpenShell framework, an open-source secure runtime designed to address security challenges in AI systems. Key partners include Cisco, which is developing DefenseClaw on OpenShell, and JFrog, which will integrate its tools to scan and verify agent skills. Other contributors are CrowdStrike, providing SafeMind for defense testing; Palo Alto Networks, offering Prisma AIRS for continuous red teaming; Capital One, developing VulnHunter for AI-powered code security; and ReversingLabs, which supplies Spectra Assure for software package analysis.

Defined security requirements and enforceable controls are critical for AI systems, with each layer of the agent stack carrying specific security responsibilities. Named owners must be assigned for security measures, and evidence of effective protections is essential. Human approval is required for consequential actions and permission changes, ensuring that access permissions do not automatically extend to data exports. Testing protocols must be established to cover attempts to alter permissions or disrupt monitoring, with ongoing evaluations necessary to address any failures. This initiative follows NVIDIA's previous discussions on open versus closed AI models, emphasizing the need for robust security frameworks in AI deployments.