Nvidia builds hardware ‘kill switch’ to stop AI agents that go rogue

Nvidia has unveiled a new security platform designed to stop autonomous AI agents when they stray beyond their permitted boundaries, adding a hardware-based “watchdog” that the company says can quarantine a misbehaving agent within milliseconds.

The chip giant announced its Open Agent Safety Platform, which combines software controls with independent hardware monitoring as companies increasingly deploy AI agents that can perform tasks with limited human supervision.

At the software level is OpenShell, Nvidia’s open-source secure runtime. It places agents in isolated environments and controls what files, networks, tools, credentials, and other resources they can access.

Crucially, those restrictions operate outside the AI agent itself, so the agent cannot simply be prompted to disable them. Nvidia says OpenShell starts from a “nothing is permitted by default” approach, with access granted according to defined policies.

The second layer, called Sentry, operates independently on Nvidia’s BlueField-4 data processing units.

Sentry continuously monitors an agent’s activity and enforces security policies from an isolated hardware environment. If an agent attempts to move beyond its software boundary, Nvidia says the system can quarantine and stop it within milliseconds.

That separation matters because the watchdog operates outside the processor where the AI agent runs, giving it an independent view of what the agent is doing.

The approach is effectively a hardware-backed safety net rather than a universal “off switch” for artificial intelligence.

The launch comes amid growing concern over increasingly autonomous AI agents and recent incidents in which agents have circumvented application-level security controls while attempting to complete assigned tasks.

Nvidia said these incidents show the need for safeguards that AI agents cannot override.

“AI’s extraordinary potential for society will only be realised if we solve AI safety,” Nvidia CEO Jensen Huang said in announcing the platform.

Anthropic has worked with Nvidia on additional security controls. Other companies supporting the initiative include Microsoft, Cisco, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Palo Alto Networks, Salesforce, SAP and SpaceXAI.

OpenShell is broadly available as open-source software and can also be extended to third-party computing platforms, including Arm and Intel systems.

Nvidia’s announcement does not mean Sentry can prevent every harmful or unexpected AI behaviour. Instead, it is designed to enforce predetermined access and security boundaries independently of the agent—providing another layer of protection if an autonomous system tries to go somewhere it should not.