Austin – October 12, 2026 -- NVIDIA introduced the NVIDIA Open Agent Safety Platform, an open source platform and reference design that applies multi-layered security controls across the software, runtime and infrastructure where AI agents operate, with CrowdStrike collaborating on the effort.
NVIDIA Sentry enforces policy from BlueField-4 DPUs and can quarantine agents in milliseconds
At the software layer, NVIDIA OpenShell provides an open source secure runtime that governs how agents execute, what they can access and where inference occurs. NVIDIA Sentry adds an independent watchdog running on NVIDIA BlueField-4 DPUs. It uses NVIDIA DOCA for out-of-band, in-silicon monitoring of agent activity and policy enforcement.
Sentry provides attested telemetry, verifies agent identity and enforces Zero Trust access to data, tools, APIs and services. During agent testing and evaluation, it can quarantine agents in milliseconds. NVIDIA Vera CPUs supply compute for agentic workloads.
CrowdStrike argues agents must be treated as privileged identities, not their own security boundary
CrowdStrike's position is that agent safety is a cybersecurity problem requiring a full-stack approach. Model safeguards are necessary but do not cover the whole security boundary once agents hold identities, credentials and access to critical systems.
The company calls for least privilege, isolation, explicit authorization, observability and enforcement outside the agent itself. Agent permissions should be bounded by policy and not expanded because an agent decides it needs more access.
Falcon platform links agent telemetry to endpoint, identity, cloud, data and AI context
Actions such as accessing a file, calling an API or opening a network connection are not inherently malicious, CrowdStrike said; context determines risk. The Falcon platform supplies that context by combining security across endpoint, identity, cloud, data and AI with telemetry, adversary intelligence and frontline expertise. Defenders and autonomous defense agents can then assess agent activity alongside other enterprise events and act on malicious behavior from one platform.
New platform extends earlier CrowdStrike-NVIDIA integrations at the runtime and infrastructure layers
CrowdStrike recently launched Falcon Guardian, extending protection across the AI attack surface, including runtime protection for AI agents. Earlier this year, the two companies unveiled a Secure-by-Design AI Blueprint that pairs Falcon protection with NVIDIA OpenShell to embed governance, visibility and control into the agent stack.
At the infrastructure layer, an existing integration feeds NVIDIA DOCA Argus telemetry into CrowdStrike Falcon Next-Gen SIEM. CrowdStrike also cites its membership in the Open Secure AI Alliance and CrowdStrike SafeMind, an agentic system for defenders built with NVIDIA Nemotron.