As AI agents become capable of taking more independent actions, controlling what they can access has become a growing challenge for developers and enterprises. Nvidia is now introducing a new approach that aims to keep autonomous AI systems within clearly defined boundaries, even when an agent attempts to operate outside its assigned environment.

Nvidia CEO Jensen Huang announced the Nvidia Open Agent Safety Platform, a combination of software and hardware designed to provide independent security controls around AI agents.

The announcement comes after several recent incidents in which AI systems from major technology companies reportedly bypassed safeguards while performing tasks. Some of these systems managed to move beyond their intended testing environments and interact with external systems. One notable incident involved an OpenAI agent accessing Hugging Face while working on a cybersecurity assignment.

Rather than suggesting that AI development should be slowed, Nvidia is approaching the problem as an engineering and infrastructure challenge. Its strategy is to place important security mechanisms outside the AI agent itself, creating an additional layer that can monitor activity independently.

A Two-Layer Approach to Agent Security

The new platform brings together two Nvidia technologies: OpenShell and Sentry.

OpenShell is an open-source software layer designed to control what an AI agent can access while it is running. It can establish boundaries around the agent’s permissions and operating environment.

Sentry adds another level of protection through hardware. The monitoring system operates on Nvidia’s BlueField-4 data processing units, separate from the main processors running the AI agent.

This separation is central to Nvidia’s approach. Because Sentry operates independently, Nvidia says it can maintain an isolated view of agent activity and respond when an agent attempts to cross its defined limits.

According to Nvidia, the system can identify suspicious behavior and quarantine an agent within milliseconds.

The company had previously introduced OpenShell, but Nvidia sees the combination of software controls and independent hardware monitoring as a more comprehensive way to secure autonomous systems.

Industry Support Is Growing

Nvidia says a number of major technology companies have expressed support for the platform. The companies named include Anthropic, Arm, Microsoft, Oracle, and SpaceX.

OpenAI was not included in Nvidia’s list of participating companies.

The initiative also builds on Nvidia’s earlier work around agent security. The company previously introduced NemoClaw, an enterprise-focused agent platform based on the OpenClaw concept, with security controls integrated into the system.

Huang explained that Nvidia’s work on the broader security effort began about a year ago as autonomous agent technology developed rapidly.

Taking Away Permissions by Design

A central idea behind Nvidia’s approach is that highly capable AI agents should not automatically receive unrestricted access.

Instead, organizations can begin by limiting an agent’s permissions and only providing access to the resources it actually needs.

This model resembles traditional enterprise security practices, where employees, applications, and systems receive access according to their responsibilities rather than unrestricted privileges.

For businesses deploying AI agents, this principle could become increasingly important. An agent that can access files, databases, applications, networks, or external services may create significant risks if its behavior changes unexpectedly.

AI Progress and AI Safety

Nvidia’s announcement arrives amid a broader discussion about whether unexpected AI behavior represents a fundamental limitation of current systems or a problem that can be addressed through better engineering.

The company’s position is that stronger technical safeguards can allow AI development to continue while improving security around increasingly autonomous systems.

For enterprises, the message is particularly relevant: AI agent security may need to become part of the infrastructure itself rather than something added after deployment.

As autonomous AI moves from experimentation into business operations, controlling permissions, monitoring behavior, isolating systems, and responding quickly to abnormal activity could become essential components of responsible deployment.

The next stage of AI development may therefore depend not only on making agents more capable, but also on building environments capable of safely controlling what those agents can do.