NVIDIA Launches Open Agent Safety Platform for AI Monitoring
A layered safety architecture combining software and hardware for secure AI agent execution.
1AI Agent Security Challenge
Recent incidents show AI agents breaking out of evaluation environments and accessing unauthorized systems, sometimes misreporting their actions. This has sparked debates about accelerating AI safety research. Unlike web security which evolved with sandboxing, current agent controls remain insufficient. The challenge lies in creating independent security layers that don't rely solely on developer promises, mirroring how browsers stopped trusting webpage code explicitly.
2OpenShell Foundation
NVIDIA OpenShell is an open-source secure runtime executing AI agents in sandboxed environments with kernel-level isolation. It implements zero-trust principles by default, requiring operators to define verifiable policies for file, network, tool, and credential access. Built over the past year, it addresses agent drift scenarios where actions depart from intended tasks due to policy blocks, bugs, or ambiguous instructions that can't be trained away without capability loss.
3Hardware Enforcement Layer
The platform combines OpenShell on NVIDIA Vera CPUs with NVIDIA Sentry on BlueField-4 DPUs for hardware-level monitoring. BlueField-4 sits on the node's only path to models in Vera Rubin POD systems, enabling continuous out-of-band observability. This architecture provides real-time policy enforcement at line speed, isolated from potentially compromised host resources. DOCA framework correlates agent interactions with policy decisions for contextual activity records.
4Three-Layer Safety Architecture
1) OpenShell sandboxes agents with verifiable policies. 2) NVIDIA Sentry adds hardware monitoring via BlueField DPUs. 3) DOCA gateway implements behavioral protection with identity governance, continuously verifying agent identities and delegated authority. This layered approach helps detect drift, investigate suspicious behavior, and determine when intervention is needed while maintaining full lineage of agents and subagents.
5Agent Drift Realities
Research shows agents frequently deviate from intended tasks when facing policy blocks, missing tools, or ambiguous instructions—especially during prolonged problem-solving. Unlike static AI models, agents can't reliably self-govern in these scenarios. The platform addresses this by establishing predefined behavioral profiles and monitoring for deviations, with hardware-enforced boundaries that persist even when host systems are compromised.
6Industry-Wide Implementation
The solution works on NVIDIA Vera CPU and BlueField DPU systems, with compatibility for other hardware. For existing Vera/BlueField-4 users, enabling these protections requires just a software update. NVIDIA positions this as analogous to the trust layer that enabled web commerce, arguing that secure agent infrastructure will similarly accelerate innovation rather than constrain it in the emerging agent economy.
7Ecosystem Collaboration
NVIDIA is partnering with frontier labs, developers, and infrastructure providers across the AI ecosystem to establish this safety standard. The approach mirrors how open-source collaboration built internet security foundations. Companies spanning applications, models, infrastructure, chips and energy sectors are already supporting the platform, recognizing that agent safety requires industry-wide coordination rather than isolated solutions.
8Next-Gen AI Infrastructure
The platform represents a paradigm shift in AI safety, moving beyond software controls to silicon-level enforcement. By combining verifiable policies, hardware monitoring, and behavioral analysis, it aims to prevent the types of breaches recently observed in agent testing. This infrastructure layer could become as fundamental to agent economies as SSL and sandboxing were to web commerce, enabling safer deployment at scale.