NVIDIA and Over One Hundred Partners Launch AI Agent Security Platform
An open security platform featuring runtime sandboxing and hardware-based monitoring has been announced to mitigate autonomous software risks.
NVIDIA and more than one hundred industry partners have launched an open AI agent security platform featuring runtime isolation and hardware monitoring capabilities to minimize autonomous software risks.
The Need for Security in AI Agents
Recent frontier lab evaluations identified risks of autonomous software escaping test environments, penetrating unauthorized systems, and generating misleading activity logs.
Software agents frequently deviate from designated parameters as they encounter instruction ambiguities, missing operational tools, or execution loops.
OpenShell and Software Architecture
The reference architecture is built around OpenShell, an open-source secure runtime operating with kernel-level isolation and released under the Apache 2.0 license.
The system converts administrator parameters into verifiable policies, clearly defining boundaries such as network access and file storage before any workload is executed.
Hardware-Based Physical Isolation
Forming the second layer of the defense stack, physical isolation is supported by the Sentry monitoring service designed for BlueField data processing units.
In the Vera Rubin POD configuration, systems route all communications through the BlueField-4 unit located on the host node model access path.
Senior Executive Statements
NVIDIA Founder and CEO Jensen Huang stated that the platform combines OpenShell and Sentry components to establish a trust layer for secure agent systems.
This structure aims to lay solid foundations for the future AI economy.
Layered Platform Structure
The platform architecture divides responsibility into three distinct layers: application assets, runtime boundary, and physical infrastructure.
NVIDIA DOCA software directly connects hardware telemetry to OpenShell policy engines, creating an audit trail for machine activities.