nemo-guardrails

Runtime safety guardrails prevent prompt injection by validating LLM inputs and outputs. Programmable rails detect jailbreak attempts and toxic content, ensuring your application only generates secure and trustworthy responses.

PII filtering and hallucination detection are handled by runtime guardrails that validate LLM outputs. You can enforce fact-checking and mask sensitive information programmatically before responses reach the user.

Yes, you can define custom LLM safety policies using the Colang 2.0 DSL. This domain-specific language allows you to program specific guardrails, workflows, and input validation rules for your application's security needs.

Runtime LLM security can run efficiently on T4 GPUs for production environments. The guardrails framework processes jailbreak detection and toxicity checks with optimized performance suitable for customer-facing applications.

The best way to add programmable guardrails is using a framework like NeMo Guardrails. It provides runtime safety mechanisms for jailbreak detection and input validation, which you can install via pip and configure using Colang.

Runtime security provides an additional layer of protection by enforcing programmable guardrails that fine-tuning cannot guarantee. It actively detects prompt injections and filters PII in real-time, preventing harmful outputs dynamically.