Защита LLM от инъекций и джейлбрейков

★ 7.7 · security

prompt-guard is a Claude Code skill that integrates Meta's Prompt-Guard-86M classifier to detect prompt injection and jailbreak attempts in LLM applications. The 86M-parameter model assigns each input one of three labels — BENIGN, INJECTION, or JAILBREAK — achieving 99%+ true positive rate with under 1% false positive rate and inference latency below 2ms on GPU. Three workflows are covered: filtering user messages before they reach the LLM, scanning third-party API responses and scraped data with a lower threshold (0.3), and batch-processing RAG document sets with a configurable batch size. Built on transformers and torch, it supports 8 languages and is straightforward to deploy via HuggingFace model hub. Teams building RAG pipelines or multi-turn chat applications who need a lightweight, fast input-validation layer without heavy infrastructure will find it a practical fit.