Nvidia‘s agentic AI stack launched at GTC this week with security built in from day one — the first major AI platform to do so, according to the announcement. Five security vendors shipped protection alongside it, four with active deployments and one with a validated early integration.
Jensen Huang framed the stakes from the keynote stage: “Agentic systems in the corporate network can access sensitive information, execute code, and communicate externally. Obviously, this can’t possibly be allowed.”
The threat context explains the timing. 48% of cybersecurity professionals now rank agentic AI as the top attack vector heading into 2026. Only 29% of organizations feel fully ready to deploy these technologies securely. Machine identities outnumber human employees 82 to 1 in the average enterprise.
IBM‘s 2026 X-Force Threat Intelligence Index documented a 44% surge in attacks exploiting public-facing applications, accelerated by AI-enabled vulnerability scanning.
Nvidia defined a unified threat model designed to flex across five vendor implementations. The company also names Google, Microsoft Security, and TrendAI as Nvidia OpenShell security collaborators.
Five Vendors, Five Enforcement Points
The five vendors map to distinct layers of the stack, not overlapping positions. CrowdStrike‘s Falcon platform embeds at four enforcement points within the Nvidia OpenShell runtime: AIDR at the prompt-response-action layer, Falcon Endpoint on DGX Spark and DGX Station hosts, Falcon Cloud Security across AI-Q Blueprint deployments, and Falcon Identity for agent privilege boundaries.
Palo Alto Networks enforces at the BlueField DPU hardware layer within Nvidia’s AI Factory validated design. JFrog governs the artifact supply chain from registry through signing. WWT validates the full stack pre-production in a live environment. Cisco runs an independent guardrail at the prompt layer.
CrowdStrike and Nvidia are also building what they describe as intent-aware controls — inline enforcement that tracks agent state across sessions, not just point-in-time inspection.
Governance Gaps That Still Exist
No single vendor covers all five governance layers. The five layers — agent decisions, local execution, cloud operations, identity, and supply chain — each carry a specific ungoverned risk if left unaddressed.
Poisoned input triggering privileged actions, agent-to-agent privilege escalation, and compromised models reaching production are among the documented failure modes the announcement maps to each layer.
The framework draws from the five vendor announcements and the OWASP Agentic Top 10. It is an analyst-synthesized reference architecture, not Nvidia‘s official canonical stack.
Security leaders are advised to treat three or more unanswered vendor questions across the five layers as a signal of ungoverned agents running in production.
Photo by Markus Winkler on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
