AI Security Is an Engineering Problem - How to Solve It at Every Layer of the Agent Stack
NVIDIA, Monday, September 21st, 2026
NVIDIA argues AI agent security requires defined controls, ownership and evidence across the model, harness and runtime layers of the agent stack.
NVIDIA frames AI security as an engineering discipline requiring defined requirements, enforceable controls, named owners and evidence that protections work.
It describes the agent stack as models providing capability, harnesses organizing context and tools, and runtimes providing execution infrastructure, each carrying its own security responsibilities.
In an example, an agent asked to export customer data after encountering malicious instructions should be blocked by network policy, with protected logs capturing the attempted call and outcome.
NVIDIA stresses that permission for one action, such as updating a record, should never automatically extend to another, such as exporting data, and that agents cannot self-authorize additional access.