XJailGuard

Project highlights

• Modular guardrail pipeline with separate input, output, and context-aware classifiers.
• Multilingual attack detection to handle unsafe prompts beyond English-only threat models.
• Token-level SHAP explanation views for blocked prompts and blocked model responses.
• Gradio-based interactive interface for security testing and model-behavior auditing.

What it is

XJailGuard is an explainable LLM safety gateway for multilingual conversational systems. It wraps a generative model with input screening, multi-turn jailbreak detection, output filtering, and token-level explanation so unsafe requests and responses are blocked before user delivery.

Problem it solves

Single-layer moderation often misses role-play jailbreaks, translated attack prompts, and chained multi-turn escalation. XJailGuard solves this by applying defense-in-depth at both ingress and egress, while exposing explainability artifacts for review and policy tuning.

How it works

• Accept user prompt and route it through multilingual input classifiers that score jailbreak intent before generation.
• Evaluate recent conversational context with a dedicated multi-turn detector to catch staged or progressive attack chains.
• Forward safe requests to a quantized Vicuna-family generation layer and collect candidate model output.
• Pass generated output through a second safety classifier to prevent unsafe responses from leaving the system.
• Generate SHAP-based token attributions whenever content is blocked so operators can inspect why a decision was made.

Architecture diagram

Pipeline showing prompt filtering, LLM generation, response validation, and explainability stages.

Key capabilities

• Modular guardrail pipeline with separate input, output, and context-aware classifiers.
• Multilingual attack detection to handle unsafe prompts beyond English-only threat models.
• Token-level SHAP explanation views for blocked prompts and blocked model responses.
• Gradio-based interactive interface for security testing and model-behavior auditing.
• Lazy LLM loading and quantized inference path for practical deployment on constrained hardware.

Impact and outcomes

• Reduces jailbreak exposure for assistant workflows by filtering threats at both prompt and response stages.
• Improves trust and triage speed by pairing each block decision with transparent token-level evidence.
• Creates an extensible blueprint for explainable guardrails in multilingual production assistants.