CrowdStrike and NVIDIA take AI agent security into silicon

At NVIDIA GTC 2026, CrowdStrike announced an expanded collaboration with NVIDIA, on October 21, to extend AI-native security from software to silicon. The move matters because it answers the question every enterprise running autonomous agents should be asking: where does enforcement actually live when the software layer is the thing being attacked?

What was announced

CrowdStrike is integrating Falcon Guardian, its AI Detection and Response product, with NVIDIA OpenShell secure runtime software, built on the OpenShell Reference Architecture. That integration connects agent prompts, model interactions, and runtime activity with broader Falcon platform security context, so malicious behavior across AI agent workflows can be detected against the same telemetry the platform already collects.

The second part is planned, not shipping: an integration with NVIDIA Sentry on BlueField-4 data processing units that would bring out-of-band monitoring and in-silicon policy enforcement to AI Detection and Response. The point of that layer is containment beyond the software runtime, so unsafe agent behavior can be detected and stopped even when the software layer misses it. Reporting puts Sentry availability in early 2027, so treat it as direction, not something to redesign around today.

Why this matters

AI agents are a new attack surface, and guardrails alone will not stop an adversary. That is CrowdStrike chief business officer Daniel Bernard's framing, and it is the right one: a prompt filter sits where the agent operates, and an attacker who steers the agent is already past it. Enforcement that lives underneath the software, in the infrastructure the agent runs on, is a different kind of control, because it does not depend on the compromised layer policing itself.

Autodesk chief trust officer Sebastian Goodwin put the enterprise version of it plainly: Falcon Guardian gives visibility into what agents are doing, what they can access, and when to intervene, and extending that control deeper into infrastructure helps customers trust the AI behind their work. That is the buying question in one sentence: can you see what your agents do, and can you stop them?

What to do

Inventory every agent running in your environment and what each one can access, including tools, data, and credentials. Require runtime visibility in procurement: prompts, tool calls, and actions taken, not just a promise that the model is safe. Scope agent permissions like human permissions, least privilege with audit trails, and log agent actions the way you log user actions. Watch Sentry availability before making hardware decisions, and do not let a planned integration carry a purchasing case that the shipping OpenShell integration has to carry on its own.

One honest caveat: CrowdStrike is one of eleven cybersecurity vendors NVIDIA named as building OpenShell support, alongside Zscaler, Okta, Palo Alto Networks, Cisco, Fortinet, Check Point, F5, 1Password, Microsoft, and TrendAI. This is an ecosystem direction, not a solo moat, and no financial terms were disclosed.