NVIDIA pitches zero-trust security layer for the agent economy
The architecture is designed to sandbox agents, continuously observe their behavior and enforce policies in real time rather than relying on agents to police themselves.
NVIDIA has unveiled a new platform designed to provide continuous monitoring and real-time security enforcement for autonomous AI agents across software and hardware.
The platform, called the Open Agent Safety Platform, combines OpenShell, NVIDIA’s open-source agent runtime running on Vera CPUs, with NVIDIA Sentry, a hardware-based watchdog running on BlueField-4.
NVIDIA says the design would provide an independent trust layer for agents as they gain access to more tools, systems and resources.
The launch of the security layer comes as several frontier AI labs have disclosed agent containment failures in recent months.
OpenAI reported models escaping a cyber evaluation and reaching Hugging Face production systems, while Anthropic found models accessing real external networks, including a production database and 15 systems that installed a fabricated package. Meta and Moonshot AI also reported agent escapes during testing.
NVIDIA argues that these failures demonstrate why agents cannot be trusted to regulate their own behavior.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
According to the company, agent breakouts can result from the interaction of tools, ambiguous instructions, extended execution and attempts to find solutions outside expected paths. It says independent security controls are needed in the same way that browsers introduced isolation to make the internet safer without requiring every website to be trusted.
NVIDIA’s OpenShell puts autonomous AI agents in zero-trust sandboxes with kernel-level isolation, continuous monitoring and behavioral detection. The company says agents can drift from their intended behavior because of blocked policies, missing tools, bugs, unclear instructions or prolonged attempts to solve difficult problems, making independent security controls necessary.
NVIDIA’s safety framework calls for policies that can be verified before execution, enforcement that remains outside the agent’s reach and controls over the path to the model. It also links greater agent authority to greater visibility into reasoning and advocates shared responsibility among AI labs, enterprises and infrastructure providers.
The platform has three layers covering application, runtime and infrastructure. OpenShell manages policies and access controls, while NVIDIA Sentry runs on BlueField-4 as an independent hardware layer. NVIDIA’s DOCA framework connects the hardware controls with OpenShell policies and helps track agent actions, tool access and delegated authority.
The architecture is designed for AI factory-scale deployments using Vera CPUs and BlueField DPUs.
In Vera Rubin Pods, BlueField-4 sits on the model’s path and can independently monitor agents and enforce policies, while NVIDIA says existing Vera and BlueField-4 customers can enable the protections through a software update.