NVIDIA OpenShell and Sentry: Policy-Enforced Security for AI Agents

Enterprises deploying large numbers of AI agents face the problem of limiting what each agent can access, rather than relying on prompts. The presentation explains NVIDIA OpenShell and NVIDIA Sentry as mechanisms for enforcing agent permissions at system-call, file, and network levels, with optional hardware-isolated monitoring.

■ Agent security approach
- Limitations of prompt-based restrictions
- Policy definition and enforcement outside the agent
- Formal verification against permission ceilings

■ OpenShell architecture
- Isolated sandbox with no direct network route
- Supervisor for outbound requests and key handling
- Gateway control plane and policy prover
- Linux kernel isolation and runtime policy enforcement

■ Sentry and BlueField-4
- Optional protection outside the host
- Hardware-isolated trust domain
- Policy updates and model connection quarantine
- Tamper-evident activity records

■ Deployment
- Open source under Apache 2.0
- Runs from edge to cloud
- Two-command adoption via CLI and sandbox creation

It is aimed at security engineers, platform teams, and AI infrastructure operators evaluating controls for agent deployments. Viewers gain an overview of how policy enforcement and isolated execution can be applied to agent access. The material focuses on architecture and deployment, not on pricing, benchmarks, or detailed implementation guidance.