prompt-guard

Detect prompt injections and jailbreak attempts in LLM prompts.

1.0k|117|Updated Feb 26, 2026

Search Tags:#security #llm #safety #input-validation #multilingual #prompt-injection #jailbreak-detection

Ready to run

npx skills add https://github.com/OpenLAIR/dr-claw --skill prompt-guard-openlair

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: prompt-guard
Source: https://github.com/OpenLAIR/dr-claw/tree/main/skills/safety-alignment/prompt-guard
Command: npx skills add https://github.com/OpenLAIR/dr-claw --skill prompt-guard-openlair

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Prompt-based systems are vulnerable to prompt injections and jailbreak attempts that can cause unsafe or undesired model behavior. This Skill provides a lightweight classifier to detect and block such inputs before they reach the model, reducing risk of data leakage and policy violations.

Core Features & Use Cases

Fast 86M parameter classifier for real-time detection.
Multilingual support across 8 languages for diverse prompts.
Plug-and-play integration with HuggingFace and RAG pipelines for secure deployments.

Quick Start

Provide a sample user prompt and observe the jailbreak score.

Frequently Asked Questions about prompt-guard

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I detect prompt injections and jailbreak attempts in real-time LLM prompts?▼

To detect prompt injections and jailbreak attempts in real-time, you can use an 86M parameter lightweight classifier that scores user inputs and API prompts before they reach the model, reducing risks of data leakage.

Can I integrate jailbreak detection into my existing HuggingFace or RAG pipelines?▼

Yes, prompt injection detection offers plug-and-play integration with HuggingFace and RAG pipelines, allowing you to secure deployments by blocking malicious inputs before they trigger unsafe model behavior.

Does prompt injection detection support multilingual inputs for LLM security?▼

Multilingual prompt injection detection supports prompts across 8 languages, allowing you to classify diverse user inputs and identify jailbreak attempts regardless of the language used.

What is the best way to prevent data leakage from LLM jailbreaks without adding latency?▼

Using a fast 86M parameter classifier for real-time scoring is the best way to prevent data leakage from jailbreaks, as it evaluates API prompts quickly and blocks policy-violating inputs pre-inference.

Are there limitations to using a lightweight classifier for prompt injection detection?▼

While a lightweight classifier provides fast real-time scoring for prompt injections, it is designed as a pre-inference filtering layer and should be combined with broader safety mechanisms for complex adversarial edge cases.