CrowdStrike and Nvidia Expand AI Agent Security Partnership

CrowdStrike is expanding its Nvidia partnership to help security teams monitor AI agents at runtime, with hardware-based enforcement planned.

At Nvidia GTC Berlin on Oct. 21, CrowdStrike announced that it is integrating Falcon Guardian, its AI Detection and Response (AIDR) solution, with Nvidia’s OpenShell runtime. The integration aims to connect agent prompts, model interactions, and runtime activity with the broader Falcon platform’s security context.

OpenShell is part of the Nvidia OpenShell Reference Architecture, which aims to provide an agentic security framework across both software and hardware. 

At the moment, CrowdStrike is pursuing integration at the software level, through OpenShell. The company says that it plans to integrate Nvidia Sentry on BlueField-4 data processing units (DPUs) to combine out-of-band monitoring with in-silicon policy enforcement to help detect and contain unsafe agent behavior.

“CrowdStrike pioneered AI-native cybersecurity. Together with Nvidia, we're setting the standard for securing autonomous AI, from detecting threats in software to stopping them at the silicon level,” said Daniel Bernard, chief business officer at CrowdStrike.

What the CrowdStrike integration brings to Nvidia OpenShell

Nvidia OpenShell is open-source software that runs AI agents within isolated environments and enforces policies governing their access to resources.

Crowdstrike’s Falcon Guardian provides the governance policies and monitoring for an agentic environment. The platform gives developers visibility into which agents are running, enforces access policies, and also provides traceability for agent actions.

By using AI and OS telemetry, Falcon Guardian is able to create an agent graph to trace the agent's runtime behavior. 

Furthermore, the Falcon Guardian platform provides security by investigating prompts to avoid threats like prompt injection and data exfiltration attacks. 

CrowdStrike’s planned integration would extend Falcon Guardian to a runtime designed to support cloud, hybrid, on-premises, edge, and air-gapped infrastructure. OpenShell already enforces access policies; Falcon Guardian would add threat detection and broader Falcon security context.

Nvidia says OpenShell’s Policy Prover uses formal verification to check whether policies stay within an allowed access boundary and whether proposed network rules introduce risky additional access.

Falcon Guardian also provides controls governing how users interact with AI tools. CrowdStrike says it can detect threats involving malicious skills, including those introduced through agent workflows.

What security teams should verify before deployment

AI agents can access sensitive data, invoke tools, and trigger system changes using permissions granted by an organization. Security teams therefore need to understand both what an agent is allowed to do and what it actually does during execution.

CrowdStrike’s announcement points toward combining that runtime visibility with enforcement deeper in the infrastructure. However, the company describes the OpenShell integration as underway and the Sentry integration as planned. Its announcement does not establish general availability for the combined offering.

Before adopting the integration, teams should confirm supported environments, release timing, licensing requirements, and which controls each component will provide. They should also establish who receives alerts, who can interrupt an agent, and how an investigation would connect a user’s prompt to subsequent tool calls and system actions.

Testing should include attempts to access unauthorized files, transmit sensitive information, or invoke an unapproved tool. The aim should be to establish whether the controls detect or block those actions, what evidence they retain, and how responders can contain a compromised workflow.

These controls should support a broader AI governance program, including narrowly scoped permissions, approval requirements for consequential actions, and clear ownership of agent identities. A runtime boundary can enforce configured restrictions, but organizations still need to decide which access is appropriate.

For security leaders, the practical measure is whether the combined controls reduce an agent’s ability to cause harm while preserving enough visibility to investigate failures. That requires testing the deployed configuration and confirming available capabilities before relying on the partnership’s broader roadmap.

Read more: As security teams build their own AI agents, least-privilege access, human approval checkpoints, and audit trails become essential to keeping automated workflows under control.