Enabling Local AI Resilience with Cisco Cloud Control and NVIDIA DGX Station

NVIDIA DGX Station for Windows brings a concrete enterprise-ready solution to local AI: combining powerful hardware connected to the applications and workflows people already use, meant for individuals or entire teams to leverage AI, locally. That creates an operating challenge: understanding which agents use the system, how their requests behave, and how to apply security policy around their actions.

Local AI is shifting from chat to action. Agentic workflows demand continuous, high-performance inference—delivered with data security and cost predictability. Local AI platforms like NVIDIA DGX Station put that compute at your desk are built on the same accelerated compute architecture used in the Cisco Secure AI Factory with NVIDIA.

NVIDIA OpenShell is the secure runtime that governs how agents execute, which resources and data they can reach, and where inference is routed. It enables organizations to run autonomous agents within boundaries they control. It delivers deterministic governance, deployment flexibility across cloud, hybrid, on-prem, edge, and air-gapped infrastructure, and a single policy layer across every model, harness, and environment. Organizations gain a single, unified policy layer to define and monitor how autonomous systems operate.

Cisco Cloud Control provides an end-to-end solution to make all AI workloads on these capable machines resilient to potential risks, enforcing a user-case-specific policy framework while also providing the required observability and tokenomics view of all the AI activity, from local AI on workstations to the AI factory running on servers across an organization.

Cisco’s solution aligns with NVIDIA’s OpenShell Reference Architecture for Safe Agents with DefenseClaw being deployed on NVIDIA DGX Station for Windows to provide discovery, telemetry and enforcement hooks into local agents.

Splunk Agent Observability (available on-prem or SaaS) gathers AI activity events and traces across all AI workloads in the enterprise and surfaces the necessary observability requirements, detecting unwanted behavior drift, malicious actions or security risks, which are mitigated by Cisco AI Defense policies. All while having a seamless experience to govern the entire AI fleet through Cisco Cloud Control.

Trusted agents on DGX Station for Windows

This layout brings the Windows desktop, agents, endpoint controls, and GPU inference into one workstation. An agent can call the local model-serving endpoint, while accessing local / external tools, AI Defense inspection, and telemetry exports use their in-built hooks. The user works directly on the NVIDIA DGX Station, and operations teams govern its activity through Cloud Control.

Common agent security and telemetry paths

The Agents group makes the application layer visible: OpenAI Codex and custom agents carry out tasks through model calls, tools, and workflows. Agents with supported DefenseClaw connectors contribute hook events and receive the configured decisions. Codex uses its selected model provider; the NVIDIA DGX Station route requires a compatible endpoint, model, and agent configuration. OpenAI’s custom provider documentation describes how the provider defines the API endpoint, authentication, and protocol.

The architecture has three operational paths:

• Inference: the agent sends a model request to its configured model serving endpoint, either on the same workstation or across the network. The application owns the workflow and any subsequent tool calls.
• Security: supported agent hooks send events to DefenseClaw. In the documented Secure Client managed profile, DefenseClaw uses Secure Client machine identity to authenticate inspection requests to Cisco AI Defense. The returned decision governs the supported action through the connector. Enforcement depends on the configured mode and connector capabilities. Secure Client managed deployment describes this relationship.
• Telemetry: instrumented applications and DefenseClaw emit agent activity, timing, and decision evidence. Supported host and GPU collectors contribute infrastructure metrics. OpenTelemetry collection and configured exporters deliver the selected signals to their observability destinations.

OpenTelemetry connects the evidence

The OpenTelemetry Collector receives, processes, and exports all AI activity. This helps connect application activity to the infrastructure serving it. Agent, device, model, and trace identifiers provide the context needed to correlate evidence across services.

For Codex, configured OpenTelemetry export adds available agent events to this evidence. DefenseClaw hooks export decisions and runtime telemetry give operators complementary views of what the agent attempted, how the workflow behaved, how many tokens were consumed locally versus via an API.

Collection and export policies determine which prompts, responses, tool arguments, and identifiers leave the endpoint. Configured redaction belongs at that boundary. GPU metrics also depend on the exporters available for the selected Windows and NVIDIA environment.

Cloud Control connects the operating workflow

Cisco Cloud Control organizes these questions into a central governance plane. Fleet provides device, site, and available infrastructure context. Tokenomics helps operators inspect AI usage and estimated cost. Agent Behavior and Agent Control bring instrumented activity and configured policy state into the investigation. Hybrid Inference Monitoring and Network Security add context about workload paths and endpoint connectivity.

Consider a team investigating a malicious agent discovered on an NVIDIA DGX Station. In Cloud Control, the operator starts with the agent and inference-host context, whether they share a workstation or run on separate devices, then examines available agent traces. Those agent traces help uncover unsanctioned behavior which does not comply with the organization’s trust policies. The investigating team gets all the information they need to determine which model is being used, what tools are called and which actions is the agent taking.

More importantly, the team can create a custom policy for that agent and push that out to all devices instantly, thereby mitigating the risk immediately, not just for that agent, but similar occurrences across their AI fleet.

Cloud Control provides the interface for that investigation. The inference runtime executes the model, while connected observability and security services supply the evidence and controls available to the operator.

Example: one rogue agent, fixed everywhere

Here’s how a team handles an unsanctioned agent on an NVIDIA DGX Station.

• Start with context. The operator sees a guardrail event on a local agent in Cloud Control.
• Follow the traces. They show which model the agent used, which tools it called, and what it did before the denial.
• Find the cause. The traces reveal why the agent attempted to break the organization’s trust policy.
• Reinforce it everywhere. The team updates a policy for that scenario and pushes it to every station.

The fix covers the same behavior anywhere in your fleet. When an agent needs more than a denial, Cloud Control can use the Cisco network to quarantine its station while the team investigates.

Proven Cisco Technology powering AI Resilience

Four complementary capabilities support this operating model:

• Splunk Agent Observability can trace and evaluate instrumented agent workflows across model, tool, and retrieval calls, helping teams investigate latency, errors, quality signals, token usage, and estimated cost.
• Cisco AI Defense helps discover AI assets, validate model and application risks, and apply configured security, privacy, and safety guardrails to supported AI application and agent traffic.
• DefenseClaw connects supported agent hooks to governance, inspection, enforcement, and audit evidence on the agent endpoint.
• Cisco Secure Client supports secure connectivity and policy-based access through its deployed modules and services, and supplies machine identity in the DefenseClaw managed profile described above.

Together, these capabilities let an operator follow a task from an AI application running on an NVIDIA DGX Station, understand the agent’s behavior and intent, apply the necessary trust policies, and review the evidence behind a security decision. Cisco Cloud Control provides the common interface for that investigation.

Some of the products and features described herein remain in varying stages of development and will be offered on a when-and-if-available basis.