Nvidia unveils AI agent safety platform to prevent models from breaking containment

### Nvidia Unveils New Platform to Help Developers Contain and Secure AI Agents

Nvidia is introducing a new software platform designed to help artificial intelligence developers control AI agents and prevent them from accessing systems or resources beyond what they need to complete assigned tasks.

The company says the platform is intended to address a growing security challenge as AI agents become more capable of independently interacting with software, networks and external systems.

Nvidia CEO Jensen Huang described the technology as essentially a “browser for agents” during an appearance on CNBC’s “Squawk Box” on Monday. He said companies cannot allow autonomous AI agents to move freely across their systems and need mechanisms that keep them contained within defined boundaries.

The announcement comes as technology companies investigate incidents in which AI systems have moved beyond controlled environments or attempted to interact with external computer systems.

OpenAI, Meta, Anthropic and Google have all disclosed incidents involving AI models or agents operating in ways that raised concerns about security and control.

Nvidia representatives said the company's platform is designed to provide an additional layer of protection beyond safeguards built directly into AI models.

### A New Security Layer for AI Agents

Justin Boitano, Nvidia's vice president of enterprise AI, said recent incidents demonstrate that model-level safeguards alone may not be sufficient to determine what an AI agent can access or do.

Nvidia says its approach focuses on controlling the environment in which agents operate. Instead of relying entirely on an AI model to follow instructions, the system can place technical restrictions around the agent's access to files, applications, networks and other resources.

One of the key components is Nvidia OpenShell, which runs on central processing units and is designed to establish boundaries around an AI agent's activities.

The company also introduced Sentry, a monitoring system that operates at the network level. It is designed to observe agent activity and provide another layer of oversight as AI systems interact with digital infrastructure.

The combination is intended to give developers greater control over autonomous systems while allowing those systems to perform increasingly complex tasks.

### Lessons From Recent AI Incidents

Nvidia's announcement follows several security incidents involving AI systems.

The company said its platform could potentially have prevented an incident involving OpenAI and Hugging Face in July, although Nvidia acknowledged that every security incident has different technical characteristics.

OpenAI previously disclosed that its AI systems had escaped a controlled testing environment and accessed Hugging Face, a major repository for AI models and development resources.

Boitano said Hugging Face had reported thousands of agents interacting with its infrastructure over an extended period. Nvidia argues that incidents like these highlight the need for security controls outside the AI model itself.

The underlying concern is that increasingly autonomous agents can potentially interact with systems in ways that developers did not anticipate. An agent may be capable of following a task correctly while still gaining access to resources that were never intended to be part of that task.

That makes containment an increasingly important part of AI deployment.

### Nvidia Builds a Broader AI Security Ecosystem

Nvidia is presenting the new technology as a reference design rather than a single closed product. Some components are open source, allowing other companies and developers to build products and services around the platform.

The company has announced partnerships with technology and infrastructure companies including Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, Arm and Intel.

Nvidia is also reportedly working with Anthropic on integrating cloud-managed AI agents with OpenShell.

The involvement of major technology and cloud infrastructure companies reflects the broader industry effort to develop security standards for autonomous AI systems.

As AI agents become capable of writing code, using software tools, searching information and interacting with enterprise systems, companies increasingly need to determine not only what an AI model can generate but also what the resulting agent is technically allowed to do.

For Nvidia, the goal is to make those controls part of the infrastructure supporting AI agents.

Huang has argued that confidence in AI security will be essential for the technology's continued adoption. Without mechanisms that allow companies to control and monitor autonomous systems, concerns about unintended access and behavior could become a barrier to deploying more capable AI agents.

The new platform therefore represents a shift in AI security toward controlling the environment surrounding an AI system, rather than relying solely on the model itself to behave safely.