CrowdStrike and NVIDIA extend AI agent security from software to silicon
CrowdStrike announced an expanded collaboration with NVIDIA on October 21, 2026, at NVIDIA GTC in Berlin, aimed at extending AI agent security from the software runtime down to enforcement in hardware. The company is integrating Falcon Guardian, its AI Detection and Response (AIDR) offering, with NVIDIA's newly announced OpenShell secure runtime software, and plans a further integration with NVIDIA Sentry for in-silicon policy enforcement.
The pitch is that guardrails inside the model or application are not enough for autonomous agents. CrowdStrike chief business officer Daniel Bernard said AI agents have created a new attack surface and that guardrails alone will not stop an adversary, framing the partnership as a standard for securing autonomous AI from software detection to silicon-level containment.
In practical terms, the OpenShell integration connects agent prompts, model interactions and runtime activity with the broader security context of the Falcon platform, so defenders can spot malicious behavior across an agent's whole workflow rather than at a single prompt boundary. That matters for enterprises deploying agents that can take actions across many internal systems.
The planned Sentry integration goes further: running on NVIDIA BlueField-4 data processing units, it would pair AIDR with out-of-band monitoring and in-silicon policy enforcement, allowing unsafe agent behavior to be contained at machine speed even if the host software runtime is compromised.
For security teams evaluating agent deployments, the announcement signals where enterprise controls are heading: runtime telemetry, cross-system context and hardware-enforced limits rather than prompt-level filters. It also gives buyers a concrete checklist item - ask agent platform vendors what enforcement exists below the application layer.
Two caveats are worth noting. The OpenShell integration was announced as available direction, while the Sentry silicon enforcement is described as planned, with no delivery date. And capabilities shown at a partner keynote should be validated in a proof-of-concept against your own agent stack before they drive procurement decisions.