NeMo Guardrails — защита LLM от атак и галлюцинаций

★ 7.7 · security

nemo-guardrails is a Claude Code skill that brings NVIDIA's NeMo Guardrails runtime safety framework to LLM applications, enabling programmable, multi-layer protection without retraining the underlying model. It covers jailbreak and prompt injection detection, input/output validation, fact-checking with retrieval, hallucination detection via consistency checking, PII masking through Presidio integration, and toxicity filtering. Safety logic is defined in Colang 2.0 DSL, a declarative language where developers specify user intent patterns and bot responses as named flows, keeping safety rules separate from application code. The skill also supports LlamaGuard for standalone moderation and is production-ready on T4 GPU. It is the right choice for teams building LLM applications that require low-latency, runtime safety checks across multiple threat categories simultaneously.