Guardrails
According to NVIDIA NeMo, guardrails are a programmable control layer around LLM applications. They block, modify, or validate user input and model output that is unsafe, off-topic, malicious, or policy-violating.
On this page
Source content checked . English translation published 2026-10-27.
In business automation
Controls around a model can validate input and output and limit actions. They are useful when text or a tool call must meet specific rules, rather than relying only on an instruction in the prompt.
Examples
These examples illustrate a possible workflow. Adapt the connections and rules to your own systems.
Business process example
An agent proposes a discount and prepares a message to a customer. Before sending, code checks that the percentage is within the allowed range and that the required approval is in place. An out-of-range proposal is stopped and sent to a manager.
Relevant tools
- OpenAI Agents SDK
- Checks can be added around an agent's input and output. Permissions for actions should also be enforced in the tools themselves.
What to check
- Blocking words alone is not an adequate safeguard. Check permissions, data, and actions outside the model as well.