What is the difference between Prompt Guard and LlamaGuard?

Prompt Guard focuses specifically on detecting technical attacks like injections and jailbreaks, while LlamaGuard is designed for content moderation, such as identifying hate speech or violence.

What is the recommended threshold for high-security applications?

For high-security environments like banking or healthcare, a lower threshold of 0.3 is recommended to maximize the True Positive Rate (TPR).

Does Prompt Guard work with languages other than English?

Yes, it is natively multilingual and supports English, French, German, Spanish, Portuguese, Italian, Hindi, and Thai.

Can I use Prompt Guard for batch processing?

Yes, the skill includes optimized workflows for batch filtering large sets of documents, making it ideal for securing RAG (Retrieval-Augmented Generation) pipelines.

How fast is Prompt Guard's inference?

It is extremely fast and lightweight, with a latency of less than 2ms on a GPU and between 50-200ms on a standard CPU.

Prompt Guard Safety Alignment

Name: Prompt Guard Safety Alignment
Author: Orchestra-Research

byOrchestra-Research

•

3,983

•

安全与测试

Protects LLM applications by detecting and filtering prompt injections and jailbreak attempts with high precision.

Prompt Guard is a specialized security skill powered by Meta’s 86M parameter classifier, designed to safeguard AI applications from malicious inputs. It provides robust defense-in-depth by identifying direct jailbreak attempts in user prompts and indirect injections hidden within third-party data or RAG documents. With sub-2ms latency and support for eight languages, it offers a lightweight yet powerful solution for maintaining the integrity and safety of LLM-based systems without significant performance overhead.

主要功能

01Identifies indirect prompt injections within RAG documents and API responses

023,983 GitHub stars

03High-performance inference with sub-2ms latency on GPU

04Multilingual support for 8 major languages including English, Spanish, and Hindi

05Detects direct jailbreak attempts and roleplay-based overrides

06Configurable detection thresholds to balance security and user friction

使用场景

01Filtering malicious user prompts before they reach an LLM backend

02Implementing a multi-layered safety architecture alongside content moderation tools

03Sanitizing third-party data and web-scraped content in RAG pipelines

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add orchestra-research/ai-research-skills prompt-guard

For use in Claude.ai and ChatGPT

Download Skill