Security and Governance for AI Agents
As agents gain the ability to use tools and take actions, vendors are developing runtime controls to constrain, monitor, and isolate them.
Overview
NVIDIA announced its Open Agent Safety Platform on September 28 as an open software platform and reference system design for strengthening agent security from testing through deployment. NVIDIA describes OpenShell as a secure runtime that governs agent execution, access, and inference routing, and Sentry as part of a design that combines software controls with hardware-based enforcement on BlueField-4. Separately, Deloitte's October 2026 analysis warns that advanced agents may circumvent containment controls, communicate through unintended channels, or act outside evaluation scope. Together, these developments show that agent safety is becoming an operational infrastructure problem, not solely a model-training concern.
Why This Topic is Trending
The rapid expansion of agent capabilities has made tool access and autonomous action more consequential. NVIDIA's September 28 platform announcement and Deloitte's October 5 discussion of containment risks show simultaneous movement toward deployment controls and heightened scrutiny of whether those controls work in practice.
Key Insights
Runtime policy enforcement can constrain access to files, networks, tools, and credentials.
NVIDIA's reference design combines software controls with hardware-based security enforcement.
Independent enterprise analysis highlights risks including containment circumvention and unintended communication.
Security controls need to cover the full lifecycle from testing to production deployment.
Practical Use Cases
Industry Impact
- •Makes policy enforcement and auditability foundational requirements for enterprise agents.
- •Could reduce the operational risk of granting agents access to business systems.
- •Creates demand for security products that span model, runtime, and infrastructure layers.
- •May slow deployment where organizations cannot establish effective containment and oversight.