NVIDIA today expanded the number of guardrails developers of artificial intelligence (AI) agents can employ to ensure content safety.
As organizations embrace agentic AI, many of them are discovering they need to address a range of governance, security and performance issues. NVIDIA has developed a set of reusable NVIDIA NIM microservices for AI guardrails that makes it simpler to embed controls that address these issues.
The latest additions to the NVIDIA NeMo Guardrails portfolio include a content safety microservice that prevents harmful output from being generated, along with a microservice that can be used to limit the range of topics that an AI agent can address. The content safety guardrail was trained using an Aegis Content Safety Dataset created by NVIDIA, which the company made publicly available on the Hugging Face platform for providing access to open source large language models (LLMs) and sharing AI application development tooling.
Additionally, there is also now a microservice that prevents jailbreaks, which is an attempt by cybercriminals to bypass AI restrictions or boundaries that an organization puts in place.
Previously, NVIDIA has made available guardrails as microservices to detect personally identifiable (PII) data and to expose larger language models to external data using retrieval augmented generation (RAG) techniques.
NVIDIA at the National Retail Federation (NRF) 2025 conference this week also revealed that the NVIDIA AI Blueprint framework for creating retail shopping assistants will incorporate NeMo Guardrails microservices.
Microservices leverage lightweight, reusable containers to make it simpler for app developers to embed these controls in an AI agent. The controls will enable organizations to operationalize AI agents faster, says Kari Briski, vice president of enterprise AI models for NVIDIA. “The agentic wave is here,” she says.
Developers of AI agents can also test the safeguards they apply using NVIDIA Garak, an open source toolkit for LLM and application vulnerability scanning developed by the NVIDIA Research team. It enables AI agent developers to identify vulnerabilities involving data leaks, prompt injections, code hallucination and jailbreaks.
The degree to which the agentic wave might be held back by lack of governance capabilities and security controls is unclear, but many organizations need to be able to comply with a raft of regulatory requirements before they can rely on an AI agent to automate a task. Otherwise, the possibility that an AI agent will generate output that violates, for example, data privacy regulations, remains high.
Regardless of organizations’ approach to agentic AI, there is a clearly a critical need to ensure governance and security controls are as simple as possible to implement. Else, they will be bolted on in a way that ultimately impedes the reliability and performance of AI agents. Customers and employees alike are not going to trust an AI agent if they don’t consider it safe to invoke.
Ultimately, just about every developer of an application is at some point going to be embedding some type of AI agent into the application. The issue that needs to be addressed sooner than later is finding the least disruptive way to ensure those AI agents behave as expected.
