AI security must be treated as an engineering problem involving defined requirements, enforceable controls, and verifiable evidence that protections work. As AI agents gain capabilities like reasoning and tool use, security responsibilities must be applied across the entire stack, including models, harnesses, and runtime environments.
Effective security for agents requires enforceable boundaries that exist independently of the agent's reasoning. This includes sandboxed execution and limits on network destinations and processes. Organizations need to establish traceable identities for agents, with credentials limited to assigned tasks and policies that require human approval for consequential actions.
NVIDIA OpenShell, an open-source secure runtime, is designed to enforce policies and provide sandboxed execution. Partners such as Cisco and JFrog are collaborating with the Open Secure AI Alliance to add governance and verification layers to agent skills. Furthermore, the industry is adopting various tools for continuous testing and incident investigation, such as CrowdStrike’s SafeMind and Palo Alto Networks’ Prisma AIRS for simulations and red teaming.
Sources
- AI Security Is an Engineering Problem — How to Solve It at Every Layer of the Agent Stack （NVIDIA Generative AI, 2026-09-21）