Crate jailguard 

JailGuard — Prompt Injection Detection

Fast, accurate prompt injection detection with a simple API. The classifier (98.40% accuracy on the in-domain test set) is embedded in the library — no external files or setup required.

Quick Start

use jailguard::{detect, is_injection};

// Simple boolean check
if is_injection("ignore previous instructions") {
    println!("Blocked!");
}

// Get detailed result with confidence score
let result = detect("What is the capital of France?");
println!("Safe: {}, Confidence: {:.1}%", !result.is_injection, result.confidence * 100.0);

Features

• 98.40% Accuracy: trained on a 17-source public dataset pipeline
• Real ML: ONNX embeddings (all-MiniLM-L6-v2) + neural classifier
• Auto-setup: ONNX model auto-downloaded on first use (~90 MB, cached)
• Simple API: is_injection(), detect(), score()

API Overview

Function	Returns	Use Case
is_injection(text)	bool	Quick yes/no check
detect(text)	DetectionOutput	Full details with confidence
score(text)	f32	Raw probability (0.0-1.0)
detect_batch(texts)	Vec<DetectionOutput>	Process multiple inputs

Re-exports

pub use embedded::detect;

pub use embedded::detect_batch;

pub use embedded::is_injection;

pub use embedded::score;

pub use embedded::DetectionOutput;

pub use embedded::RiskLevel;

Modules

c_api

C ABI surface — Go (cgo) and Node.js (napi-rs) bindings link against these extern "C" functions. Compiled unconditionally so the cdylib / staticlib artifact always exposes the symbols; the c-api feature only gates the cbindgen header regeneration in build.rs. C ABI for JailGuard.

embedded

Zero-configuration embedded detector

Enums

Error

Main error type for jailguard operations.

Functions

download_model

Ensure the ONNX model is available locally, downloading it if necessary.

model_cache_dir

Public-API variant — returns the cache directory as a String.

Type Aliases

Result

Result type alias for jailguard operations.