NVIDIA Launches Open Agent Safety Platform to Put Limits on AI Agents

There’s a problem growing quietly inside organisations that are rushing to deploy AI agents: nobody’s entirely sure what those agents are doing when nobody’s watching. NVIDIA thinks it has an answer.

On 28 September 2026, NVIDIA announced the NVIDIA Open Agent Safety Platform — a combination of software controls and infrastructure-level monitoring designed to keep autonomous AI agents within strictly defined boundaries. The announcement came alongside comments from NVIDIA founder and chief executive Jensen Huang on CNBC’s Squawk Box, and it’s already drawing attention from organisations that are grappling with how to deploy AI agents responsibly.

What the Platform Actually Does

At the centre of the announcement are two components: NVIDIA OpenShell and NVIDIA Sentry.

OpenShell is an open-source runtime — reportedly at version 0.1.0, though the precise current release status should be confirmed against NVIDIA’s technical documentation before any deployment decision. Think of it as a permission layer that sits around an AI agent and converts what that agent is allowed to do into enforceable rules. Files, tools, networks, processes, credentials — OpenShell is designed to control access to all of them. It’s a sandboxed environment, meaning the agent operates inside a defined space rather than roaming freely across a system.

Sentry takes a different approach. It’s a reference design for independent infrastructure-level monitoring — hardware that watches what the agent is doing from outside the agent itself. The idea is that if an agent tries to exceed its permitted authority, Sentry can intervene. That independence matters. A monitoring system that runs inside the agent it’s monitoring is a bit like asking someone to mark their own homework.

NVIDIA says the platform covers the full journey from agent testing through to live deployment, across software, hardware, compute and robotics systems.

Huang’s Core Argument

Jensen Huang, speaking on Squawk Box, put the principle plainly. Huang said: “AI agents should have the minimum rights necessary to perform their assigned tasks.”

That’s the least-privilege principle — a concept well established in traditional cybersecurity — applied directly to AI. The idea is that an agent handling customer service queries shouldn’t also have access to financial databases or system-level processes. Restricting authority at the outset limits the damage if something goes wrong.

NVIDIA frames AI safety here as an engineering and systems-control challenge. Rather than relying on instructions or behavioural prompts to keep an agent in check — essentially trusting the agent to behave — the platform is designed to enforce restrictions at the environment level. The agent can’t exceed its permissions because the system won’t let it, not because it’s been asked nicely.

Built With Industry, Not Just by NVIDIA

The platform wasn’t developed in isolation. NVIDIA says more than 100 industry partners were involved in building it — a broad coalition that suggests the company is positioning OpenShell and Sentry as infrastructure that others can build on and adapt.

OpenShell being open-source is relevant here. Organisations can inspect the code, understand exactly how it works, and adapt it to their own environments. That’s a meaningful difference from a closed black-box solution, especially for regulated sectors where accountability and transparency matter.

But it’s worth being clear about what NVIDIA’s announcement does and doesn’t establish. The claims about preventing rogue-agent behaviour are vendor claims. The platform has not been independently certified by UK regulators, and its suitability for any specific regulated deployment — healthcare, financial services, public sector — would need to be assessed against relevant legal and professional standards. Technical controls don’t replace human accountability, risk assessment, or sector-specific regulation.

The Wider Context

Autonomous AI agents are already being used to write code, process documents, manage workflows, and interact with customers. The concern — increasingly voiced by security researchers and governance experts — is that many of these deployments lack adequate controls over what agents can access and what they can do if something goes wrong.

NVIDIA’s platform is a direct response to that concern. Whether it becomes a widely adopted standard will depend on how well it works in practice, how organisations configure and maintain it, and whether independent security assessments back up the company’s claims. The involvement of over 100 partners suggests momentum, but real-world deployment evidence will matter more than launch announcements.

The UK government’s current approach to AI regulation relies on existing regulators applying cross-sector principles — safety, security and robustness among them — rather than a single new AI law. That means organisations deploying agents in regulated UK sectors are already expected to manage these risks, with or without a platform like this one.

What This Means for Kent Residents

No specific deployment of NVIDIA OpenShell in Kent has been confirmed, but the platform is directly relevant to the kinds of organisations that serve Kent residents every day. Public bodies like Kent County Council, Medway Council, and NHS Kent and Medway ICB are all potential users of autonomous AI agents for administration, customer services, and clinical support — and any such deployment would need exactly the kind of permission controls and monitoring that NVIDIA is describing. For residents, the practical implication is indirect: better safety infrastructure around AI agents means less risk of automated systems making unauthorised decisions about personal data or public services. No action is needed from consumers right now, but this is the kind of behind-the-scenes infrastructure that shapes how safely AI enters everyday public life.