Nvidia addresses AI agent safety and security with new NeMo Guardrails NIMs

As the use of agent AI continues to grow, so does the need for safety and security.

Today, Nvidia announced a series of updates to it NeMo's guardrails technology designed specifically to address agent AI needs. The concept behind guardrails is to provide some form of policy and control for large language models (LLMs) to help prevent unauthorized and unintended results. The concept of guardrails has been adopted by several retailers in recent years, including AWS.

The new NeMo Guardrails updates from Nvidia are designed to make it easier for organizations to deploy and provide more granular controls. NeMo Guardrails are now available as a NIM (Nvidia Inference Microservices), which are optimized for Nvidia GPUs. In addition, there are three new NIM-specific services that businesses can use for content safety, theme control and jailbreak detection. The guardrails have been optimized for agent AI use, rather than just single LLMs.

“It's not just about building a model anymore,” said Kari Briski, vice president for enterprise AI models, software and services at Nvidia, in a press release. “It's about guardrails and a whole system.”

What the NeMo Guardrails bring to the Agent AI campaign

Using an AI agent it is expected to be a big trend in 2025.

Although agent AI has plenty of benefits, there are also new challenges, especially in terms of security, data privacy and regulatory requirements, which could create significant barriers to use.

The three new NeMo Guardrails NIMs are intended to help solve some of these challenges. These include:

• NIM Content Safety: Trained on Nvidia's Aegis content safety database with 35,000 human-annotated samples, this service blocks harmful, toxic and unethical content.
• NIM Topic Control: Helps ensure AI interactions stay within pre-defined topic boundaries, preventing chat drift and unauthorized information disclosure.
• NIM Jailbreak Detector: Helps prevent security bypasses through clever hacks, receiving training data from 17,000 known successful jailbreaks.

The complexity of AI defense systems

The complexity of AI defense systems is important, as they can include multiple agents and interconnected modules.

Bankruptcy we provided an example of a retail customer service agent position. Consider a person who interacts with at least three agents, a reasoning LLM, an augmented generation agent (RAG) and a customer service assistant. All of them are required to enable the live proxy.

“Depending on the user interaction, there can be many LLMs or interactions, and you have to protect each of them,” Briski said.

Although there are complexities, she noted that a primary goal with NeMo Guardrails is to make NIMs easier for enterprises. As part of today's release, Nvidia is also providing blueprints to show how the various NIM protection rails can be used for various situations, including customer service and sales.

How Nvidia's guardrails affect agent AI performance

Another major concern for enterprises using agent AI is performance.

Briski said that as enterprises use agent AI, there may be concerns about introducing latency by adding guardrails.

“I think that as people were initially trying to add guardrails in the past, they were putting more LLMs in place to try and add guardrails,” she explained.

The latest NeMo Guardrail NIMs have been updated and updated to address latency concerns. Nvidia's early tests show that protection groups can get 50% better with protection rails, which only add about half a second of time.

“This is very important when using agents, because as we know, it is not just one agent, there are several agents that can be inside an agent system,” said Briski.

Nvidia NeMo Guardrails NIMs for agent AI are available under the Nvidia AI enterprise license, which currently costs $4,500 per GPU per year. Developers can try them for free under an open source license, as well build.nvidia.com.