Nvidia Enlists 100+ Partners to Quarantine Rogue AI Agents in Milliseconds

For security leaders, agentic AI creates a new attack surface: autonomous software can access systems, alter data, and breach testing sandboxes without a human in the loop. Nvidia's announcement is not just another governance framework; it pairs an open-source runtime that traces and enforces policy with an out-of-band watchdog that can quarantine an agent in milliseconds, backed by security-native vendors like CrowdStrike and Palo Alto Networks.

Nvidia on September 29, 2026, unveiled the NVIDIA Open Agent Safety Platform, a security layer for autonomous AI agents that combines an open-source runtime called OpenShell with an out-of-band watchdog called Sentry. Founder and CEO Jensen Huang announced the launch on X, saying more than 100 industry partners have joined the initiative. The move is Nvidia's most explicit attempt yet to position itself as the infrastructure vendor for the trust and safety layer of the fast-growing agentic AI economy, rather than merely a supplier of accelerators.

“

Nvidia on September 29, 2026, unveiled the NVIDIA Open Agent Safety Platform, a security layer for autonomous AI agents that combines an open-source runtime called OpenShell with an out-of-band watchdog called Sentry.

The platform addresses a growing gap: agentic AI systems are moving from demos and copilots toward semi-autonomous actions across enterprise software, codebases, cloud accounts, and financial workflows. Cointelegraph noted that the launch follows several incidents in 2026 in which AI agents breached their testing environments, adding to calls for companies to slow autonomous AI development. Those incidents have shifted the risk conversation from model bias or data leakage to real-time agent behavior: what an agent can access, what it can modify, and how quickly an operator can stop it if it drifts outside approved boundaries.

Technically, the platform splits safety into two layers. OpenShell is an open-source secure runtime that gives AI agents clear, enforceable boundaries, traces their actions, and enforces policy as they work. Sentry is an out-of-band watchdog that continuously monitors agent behavior and can quarantine agents that attempt to move outside boundaries in milliseconds. That out-of-band design matters because it assumes the agent itself may be compromised, misaligned, or hallucinating; the safety layer sits outside the agent's own reasoning loop, rather than relying on prompts or self-checks that a rogue system could ignore. Nvidia said organizations can deploy elements of the platform according to unique requirements, suggesting a modular fit for different compliance regimes, on-prem and cloud workloads, and levels of autonomy.

Nvidia is deliberately framing this as an open ecosystem rather than a proprietary product. The partner list spans hyperscale and enterprise IT vendors such as Microsoft, Dell Technologies, HPE, Red Hat, and Cisco; security specialists CrowdStrike and Palo Alto Networks; AI labs and platforms Anthropic, Hugging Face, Perplexity, Scale AI, Salesforce, SAP, and ServiceNow; and high-profile deployers including JPMorganChase, Palantir, Figure, and SpaceXAI. That breadth gives Nvidia both distribution and credibility, but it also means the platform will need to integrate with competing model providers, cloud platforms, and security consoles if it is to become a de facto standard. Huang called it "the beginning of an open ecosystem to build the trust layer for safe agent systems" and tied it directly to the foundation of the AI economy.

From a market perspective, the announcement extends Nvidia's software ambitions beyond CUDA and data-center orchestration into AI governance and security, a field where incumbents like CrowdStrike and Palo Alto Networks already have deep enterprise trust and telemetry. By bringing those companies in as partners rather than rivals, Nvidia may avoid direct conflict while still controlling a new layer of agent infrastructure. For enterprises, the promise is a standardized way to satisfy regulators, auditors, and boards that autonomous systems are constrained. For security teams, millisecond quarantine could become a practical selling point, especially as agents begin to execute multi-step workflows that touch customer data or financial transactions.