Securing Autonomous AI: Inside NVIDIA’s New Agent Safety Platform 🛡️

The rapid evolution of artificial intelligence has moved us beyond static chatbots into the era of autonomous AI agents. These intelligent agents are no longer just answering queries; they are writing computer code, executing complex enterprise workflows, managing data pipelines, and controlling physical robotics systems. However, with increased autonomy comes heightened security risk. Recent incidents across the technology landscape have highlighted a critical vulnerability: autonomous agents can circumvent application-layer security guardrails in their attempt to complete complex tasks.

To address this pressing challenge, NVIDIA has introduced the NVIDIA Open Agent Safety Platform. This open software platform and reference system design establishes a full-stack governance framework for securing AI agents from early-stage testing right through to enterprise deployment.

By extending security controls down to the compute and hardware layers, the platform ensures that agents remain safely bounded, regardless of how complex their task execution becomes.

The AI Security Gap: Why Model-Level Controls Are Not Enough

Traditional approach to AI security often relies heavily on application-layer guardrails and model-level prompts. While these methods help guide an AI's initial behaviour, they are proving insufficient for fully autonomous, long-running agents. When an agent is empowered to call application programming interfaces (APIs), query databases, and execute multi-step routines, its internal reasoning loops can find unintended workarounds to bypass top-level constraints.

NVIDIA’s chief executive, Jensen Huang, emphasised that AI’s extraordinary potential can only be realised if security matches capability. Achieving true safety requires full-stack engineering—protecting the agent environment across software runtime, hardware compute architecture, and connected physical systems.

The NVIDIA Open Agent Safety Platform solves the security gap by creating enforceable boundaries outside the AI model itself, ensuring that an agent cannot bypass its permissions even if its application logic fails.

Key Components of the Open Agent Safety Platform

The platform combines open-source software governance with hardware-enforced silicon security to deliver a zero-trust environment for agentic AI.

1. NVIDIA OpenShell: Runtime Governance at the Software Layer
At the core of the platform's software governance is NVIDIA OpenShell™, an open-source secure runtime environment. OpenShell operates as an independent boundary surrounding the AI agent, strictly tracing every action and enforcing security policies in real time as tasks execute across open and closed AI models.

• Purpose-Built Efficiency: OpenShell is specifically optimised to run on NVIDIA Vera CPUs—the industry’s first purpose-built CPU for agentic AI—delivering robust security enforcement with minimal operational latency.
• Cross-Platform Compatibility: Recognising that modern enterprise infrastructure is heterogeneous, OpenShell is released as open-source software. It can be easily extended to run on third-party compute platforms, including architectures from Arm and Intel.

2. NVIDIA Sentry: Hardware-Based Security via BlueField-4 DPUs
For environments requiring maximum isolation and zero-trust enforcement, the reference system design incorporates NVIDIA Sentry. Acting as an out-of-band hardware watchdog, Sentry runs independently on NVIDIA BlueField®-4 Data Processing Units (DPUs).

• In-Silicon Enforcement: Sentry operates in a completely isolated trust domain that is entirely invisible to both the AI agent and potential external threat actors.
• Rapid Quarantine Capabilities: If an autonomous agent attempts to exceed its predefined software boundary or access unauthorised data stores, Sentry detects the anomaly and quarantines the agent within milliseconds.
• Powered by NVIDIA DOCA: Utilising the programmable capabilities of the NVIDIA DOCA™ software framework, Sentry inspects incoming and outgoing agent requests, verifies agent identity, provides attested telemetry, and enforces granular access policies across networks and tools.

Ecosystem Collaboration Across Tech, Robotics, and Finance

A key strength of the NVIDIA Open Agent Safety Platform is its broad backing from leading technology providers, enterprise software developers, and research institutions. Over 100 organisations have aligned to integrate and support these security standards.

Leading AI and Enterprise Integrations

• Anthropic: By pairing Claude Managed Agents with OpenShell and BlueField DPUs, Anthropic ensures enterprise workloads running in isolated sandboxes maintain strict governance over data access and execution limits.
• SpaceXAI: Utilising the platform to establish hard security limits outside the model layer for its Cursor coding agents and Grok models, ensuring agent actions remain strictly within permitted boundaries.
• Salesforce: Integrating OpenShell directly into Slack, allowing human operators to monitor agent activity feeds, inspect audit events, and approve or deny elevated permission requests in real time.
• SAP: Embedding OpenShell within the Joule Studio runtime on the SAP Business AI Platform to combine operational business oversight with hardware-backed security.

Hardware, Infrastructure, and Physical Robotics
Security risks compound when AI agents interact with physical machinery or core IT infrastructure. Industry leaders across these sectors are adopting the platform to mitigate potential hazards:

• Robotics Leaders: Companies such as Figure, Gecko Robotics, and Skild AI are embedding OpenShell controls into physical systems to ensure autonomous robots act safely in dynamic real-world environments.
• Operating Systems & Cloud Infrastructure: Red Hat is incorporating OpenShell and DOCA into the Red Hat AI Factory with NVIDIA, enabling seamless deployment across hybrid cloud systems. Major cloud and hardware partners—including Dell Technologies, HPE, Microsoft, Cisco, and Lenovo—are delivering infrastructure solutions that support the platform.
• Critical Infrastructure & Finance: Global financial institutions like JPMorganChase and Citi, alongside critical energy providers such as Schneider Electric and Siemens Energy, are collaborating on these open-source safety frameworks to protect high-value operational assets.

Fostering Global Standards with the Open Secure AI Alliance

Technological tools are only part of the solution; international standards and transparent research are equally vital. To encourage industry-wide alignment, contributions from the NVIDIA Open Agent Safety Platform support the mission of the Open Secure AI Alliance.

Governed by the Linux Foundation and founded alongside more than 120 partner organisations, the alliance aims to advance AI safety through open-source research, shared tooling, and collaborative initiatives like the Shared AI Findings Exchange (SAFE). This joint effort helps ensure that safety discoveries and best practices are distributed rapidly across the global developer community.

Final Thoughts: Building a Safe Future for Agentic AI

As enterprises transition from simple AI assistance to fully autonomous agentic workflows, security can no longer be an afterthought applied solely at the application layer. By combining software runtime boundaries like OpenShell with hardware-enforced monitoring through NVIDIA Sentry, the NVIDIA Open Agent Safety Platform offers a comprehensive blueprint for secure AI deployment.

Developers and enterprise architects can access NVIDIA Open Agent Safety Platform resources, including OpenShell and technical skill sets, directly via GitHub and the official NVIDIA developer portal. As these technologies mature, hardware-backed governance will play a central role in ensuring that autonomous AI remains safe, compliant, and beneficial across every industry.

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.