NVIDIA has launched an Open Agent Safety Platform designed to give companies greater control over autonomous AI agents, from testing through deployment. The platform extends security controls beyond the AI model itself, reaching into the software, compute, and hardware layers that actually execute agent actions.
The platform combines two core components: NVIDIA OpenShell, an open-source secure runtime for AI agents, and NVIDIA Sentry, a reference system design that uses NVIDIA BlueField-4 data processing units, or DPUs, to monitor and enforce security policies independently of the host system.
This distinction matters as AI agents move beyond answering questions and start using tools, accessing data, writing code, and taking actions across enterprise systems. NVIDIA says the platform aims to establish enforceable boundaries around those actions, rather than relying solely on the model or the agent’s own instructions.
NVIDIA moves agent security beyond the model.
Traditional AI security controls often sit around the application or model layer. NVIDIA argues that this approach can leave gaps when autonomous agents operate across multiple systems at once.
Its new architecture places security controls at several levels simultaneously. OpenShell creates a secure runtime boundary around an agent. It can define which files, networks, tools, processes, and credentials an agent can access, then enforce those policies while the agent operates. NVIDIA describes OpenShell as open-source software that uses sandboxing and kernel-level isolation to do this.
Sentry adds an independent hardware-based layer on top. Running on BlueField-4 DPUs, it continuously monitors agent activity and can enforce policies outside the agent’s own execution environment. NVIDIA says it can quarantine an agent that attempts to move beyond its permitted boundary within milliseconds.
This creates a layered architecture, rather than placing all responsibility on the model alone.
OpenShell creates a runtime boundary for AI agents.
OpenShell is the software component of the platform and is now broadly available, according to NVIDIA. The runtime lets operators define permissions before an agent begins working. Those permissions can cover access to files, networks, tools, processes, and credentials.
The system then applies those restrictions while the agent executes its assigned task. NVIDIA says OpenShell works with both open and closed AI models, giving organizations flexibility in how they deploy it.
The software is optimized for NVIDIA Vera CPUs, but since it’s open source, NVIDIA says it can also extend to third-party compute platforms, including those based on Arm and Intel. That interoperability could matter for organizations that don’t want their agent security controls tied entirely to one hardware environment.
Sentry adds a hardware watchdog.
The more unusual part of NVIDIA’s architecture is Sentry. It’s a reference system design for an out-of-band watchdog running on NVIDIA BlueField-4 DPUs. Rather than relying solely on the host environment where the agent operates, Sentry provides an independent monitoring and enforcement layer instead.
NVIDIA says Sentry can inspect agent requests and responses, verify agent identity, provide telemetry, and enforce granular zero-trust policies governing access to data, tools, APIs, and services. An agent that attempts to move outside its defined software boundary can be quarantined and stopped within milliseconds, according to the company.
This architecture aims to make the security layer harder for an agent to bypass, since enforcement happens outside the agent’s own execution environment. NVIDIA’s technical documentation describes this as an out-of-band enforcement approach, with BlueField-4 positioned as an independent infrastructure security layer.
Why agent security is becoming a bigger issue
AI agents differ from conventional chatbots because they can take actions, rather than simply generate responses. An agent may receive a goal, write or execute code, call external tools, access files, interact with APIs, and keep working as new information becomes available.
That creates a genuinely different security problem. If an agent has excessive permissions, an error, compromised instruction, or unexpected behavior could potentially affect systems well beyond the original task.
NVIDIA’s technical team describes agent drift as situations where an agent’s actions depart from its intended task or operating constraints. The company says drift can occur because of policy blocks, bugs, missing tools, ambiguous instructions, or prolonged autonomous operation. The platform therefore focuses on enforcing boundaries independently of the model itself.
More than 100 organizations are working with NVIDIA.
NVIDIA says more than 100 organizations are working with its Open Agent Safety Platform technologies. This group spans AI developers, cloud and infrastructure providers, cybersecurity companies, enterprise software firms, financial institutions, and robotics companies.
Participants named by NVIDIA include Anthropic, Cisco, CrowdStrike, Dell Technologies, Figure, HPE, Hugging Face, JPMorgan Chase, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI, ServiceNow, and SpaceXAI.
The breadth of this group reflects how agent security is becoming an infrastructure issue, rather than simply a model-development problem. NVIDIA says companies across financial services, energy, robotics, and enterprise software are exploring the technology.
Anthropic, Salesforce, and SAP add enterprise integrations.
NVIDIA says Anthropic is working with it to add security controls around Claude Managed Agents. Anthropic’s approach separates the agent loop from the sandboxes in which the agent performs its work, and NVIDIA says OpenShell and BlueField integrations can provide additional controls over access to those sandboxes.
Salesforce and NVIDIA have also integrated OpenShell with Slack. This integration lets teams view agent activity and audit events, then approve or reject requests for additional permissions, adding a human-oversight layer to agent operations.
SAP is embedding OpenShell with its Joule Studio runtime as part of the SAP Business AI Platform. SAP is also contributing engineering work to OpenShell and collaborating on interoperability standards through the Open Secure AI Alliance.
These integrations show that agent governance is being positioned as part of enterprise infrastructure, rather than as a standalone security product.
Financial services are among the early participants.
The financial sector has particular reasons to focus on agent controls. AI agents could eventually interact with customer information, internal systems, financial data, APIs, and other sensitive resources.
NVIDIA identifies Citi and JPMorgan Chase among financial-services organizations collaborating on open-source agent safety technologies. For banks and other regulated institutions, the ability to define and audit exactly what an AI agent can access could become an important part of deploying autonomous systems safely.
This technology doesn’t remove the need for conventional cybersecurity, access controls, or governance. Instead, NVIDIA positions it as another enforcement layer for systems where AI agents can act autonomously.
Robotics extends the problem into the physical world.
The security challenge becomes even more consequential when AI agents control physical systems. NVIDIA says robotics companies, including Figure, Gecko Robotics, and Skild AI, are building with OpenShell to add agent safety controls to autonomous systems.
A software agent that exceeds a permission boundary may create a data or cybersecurity incident. An autonomous system acting in the physical world could create a very different class of operational risk instead.
This is why NVIDIA’s platform extends its architecture across software, compute, and robotics alike. The objective is to establish a common control layer regardless of whether an agent operates a coding environment, an enterprise application, or a physical machine.
NVIDIA is building an open agent-security ecosystem.
NVIDIA says OpenShell is available as open-source software and can be accessed through its developer resources and GitHub. The company is also linking the work to the Open Secure AI Alliance, which NVIDIA says was initiated with more than 120 organizations and is governed by the Linux Foundation.
This broader strategy matters. Rather than creating a security mechanism exclusively for NVIDIA’s own AI models, NVIDIA is proposing an open layer that can sit around different models and infrastructure. That could let enterprises establish consistent policies even when their AI environments contain models from multiple providers.
The shift from AI guardrails to enforced boundaries
The larger development behind NVIDIA’s announcement is a change in how companies approach agent security. As AI systems become more autonomous, simply telling an agent what it shouldn’t do may not be enough. Organizations increasingly need technical controls that can enforce those restrictions independently of the agent itself.
NVIDIA’s architecture attempts to do that at two levels. OpenShell establishes the runtime boundary, while Sentry adds an independent hardware enforcement layer on top.
This approach doesn’t guarantee that an AI agent will behave safely in every situation. It also doesn’t replace broader application security, identity management, monitoring, or human oversight. But it does provide another layer between an autonomous AI system and the resources it can access.
For enterprises preparing to give AI agents greater authority, that distinction could become increasingly important. NVIDIA’s message is clear: as AI agents gain the ability to act, security controls need to govern not just what models generate, but also what those agents are technically allowed to do.
NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment


Comments
No comments yet. Be the first to share your thoughts.
Leave a comment