CrowdStrike and Nvidia partner to secure AI agents at the hardware level

Securing autonomous AI agents

As organizations begin to deploy AI agents that can perform tasks independently, security teams face a new challenge in monitoring these automated workflows. These agents can access sensitive data, invoke various software tools, and trigger significant system changes based on the permissions granted by a company. Because they operate autonomously, it is difficult for humans to keep track of every action an agent takes in real time.

CrowdStrike and Nvidia are expanding their partnership to address this specific risk by creating a framework for agentic security. The goal is to provide security teams with the ability to monitor and enforce behavior across different environments, including cloud, hybrid, on-premises, edge, and air-gapped infrastructures. By connecting agent prompts and model interactions with broader security context, the companies hope to create a standardized way to manage these tools.

The collaboration focuses on ensuring that AI agents stay within their intended boundaries while still being useful to the users. Security teams must clarify the gap between an agent's authorized permissions and its actual behavior during execution. This partnership aims to provide the visibility necessary to detect malicious skills or unauthorized actions that might be introduced through complex agent workflows.

How the software integration works

The first step of this collaboration involves integrating CrowdStrike’s Falcon Guardian with Nvidia’s OpenShell runtime. OpenShell is an open-source software designed to run AI agents within isolated environments where specific policies govern their access to resources. By using AI and OS telemetry, Falcon Guardian creates an agent graph that allows developers to trace the runtime behavior of these agents.

Falcon Guardian provides the governance policies and monitoring for the environment, while OpenShell handles the isolation. This combination gives teams visibility into which agents are running and provides a clear audit trail for their actions. It also includes features to investigate prompts to prevent common threats like prompt injection, where a user tries to trick the AI into performing unauthorized tasks.

Nvidia also contributes a tool called Policy Prover to the OpenShell architecture. This tool uses formal verification, a method of mathematical proof, to check whether policies stay within an allowed access boundary. It specifically looks for instances where proposed network rules might introduce risky additional access that could be exploited by an agent or an attacker.

Moving toward hardware enforcement

While the current integration focuses on the software level, CrowdStrike has plans to move deeper into the physical infrastructure. The company intends to integrate Nvidia Sentry on BlueField-4 data processing units, which are specialized hardware components. This move is designed to merge external monitoring with hardware-level rule application to contain unsafe behavior.

In-silicon enforcement means that the security rules are applied directly at the hardware level rather than just within the operating system. This provides a more robust defense because it can stop unauthorized actions before they reach the software layer. By using the BlueField-4 DPUs, the companies aim to create a security framework that covers both the software and the hardware components of an AI system.

Daniel Bernard, chief business officer at CrowdStrike, said the companies are setting a standard for securing autonomous AI. This progression from software monitoring to hardware-based blocking represents a significant shift in how enterprise AI security is architected.

Deployment and testing requirements

The announcement does not establish general availability for the combined offering, and the Sentry integration is still in the planning phase. Before adopting these tools, security teams will need to confirm supported environments, release timing, and licensing requirements. They must also determine which specific controls each component will provide to ensure there are no gaps in their security posture.

Organizations should establish clear protocols for who receives alerts and who has the authority to interrupt an agent if it behaves unexpectedly. Testing is essential to see how the controls react to attempts to access unauthorized files or transmit sensitive information. The goal of these tests is to confirm whether the system detects or blocks the action and what evidence is retained for investigators.

Ultimately, these controls are meant to support a broader AI governance program. This includes establishing narrowly scoped permissions, requiring human approval for consequential actions, and maintaining clear ownership of agent identities. While a runtime boundary can enforce restrictions, the organization must still decide which specific levels of access are appropriate for their specific business needs.