Open Guard Layer (OGL-Mini) 🔒 Hybrid AI Agent Security for Typescript

Open Guard Layer (OGL-Mini) - a full-featured, open-source, fast and lightweight security model (ONNX) for AI agents with ready-to-use TypeScript implementation. Works in Node 18+ and browsers (WASM).

Have a questions? Contact me

About | Installation | Get Started | More Information | Contacts

About

The OGL-Mini (Hybrid AI Guard) a full-featured solution with 3 layers of scanning and sanitizing AI agent inputs. It can be helpful to detect modern AI agents attacks (OWASP LLM Top 10 2025-2026 + Agentic Top 10).

❓ Why OGL-Mini?

🔹 Ready-to-Use Lightweight module for TypeScript; 🔹 Extra lightweight AI input security model (~300MB, ONNX runtime). Inference starts at CPU; 🔹 Support both platforms: NodeJS and Modern Browsers; 🔹 Extremely fast (great for realtime ai agents and applications); 🔹 OWASP LLM Top 10 2025-2026 + Agentic Top 10 security detector; 🔹 Production ready with benchmarks;

🔒 What kind of attacks does it protect against?

• General Attacks: LLM01 Prompt Injection + LLM02 Sensitive Disclosure, LLM03 Supply Chain, LLM04 Data Poisoning, LLM05 Improper Output, LLM06 Excessive Agency, LLM07 Misinformation, LLM08 Hidden Context, LLM09 Vector Weakness, LLM10 Unbounded Consumption + Agentic (Goal Hijack, Privilege Abuse, Code Exec, InterAgent, Tool Misuse, Memory Poisoning, Cascading, Rogue, Policy Puppetry, EchoLeak, Lies-in-the-Loop)
• Modern Injections: S3 encoding (base64), homoglyph, zero-width (\u200b), spaced letters, control tokens (<|im_start|>, [INST]), indirect JSON/tool, HTML markdown, agent-specific, best-of-n, typoglycemia - all in training (14k modern obfuscation, bilingual)
• Auto-sanitizing: Sanitize fields by length, bidi-characters, control characters, excessive whitespaces etc.

Read more about model at Repository homepage

Install

You can install Hybrid AI Guard (OGL-Mini) using NPM or directly from GitHub.

From NPM:

npm install hybrid-ai-guard
npm install onnxruntime-web onnxruntime-node  # Need to support ONNX Models

Note: the NPM package doesn't include a model weights. Please, download models from GitHub or Huggingface

From GitHub

git clone https://github.com/devsdaddy/ogl-mini/
cd ./ogl-mini/clients/typescript/
npm install && npm install onnxruntime-web onnxruntime-node

Now you can use OGL-Mini in your projects.

Get started

Quick start without OGL-Mini Model (lightweight 18-feature classifier, 0.01ms)

You can use only HybridGuard with heuristics checks and input sanitizing. This approach doesn't require to connect our onnx-based model, but it's not a full-featured scanner without inference.

Simple example without model loading:

import { HybridGuard } from "hybrid-ai-guard";

// Create our guard
const guard = new HybridGuard(); // threshold 0.60
const r = await guard.checkInput("Ignore previous instructions");
console.log(r.safe, r.label, r.riskScore, r.stage); // false direct_prompt_injection 0.97

// sync fast path (no async ONNX)
const r2 = guard.checkInputSync("Привет как дела");
console.log(r2.safe); // true

// PII
const { entities, redacted } = guard.detectPii("Иван Петров [email protected] +7 999 123-45-67");
console.log(entities, redacted);

// output guard
const out = await guard.checkOutput("My email is [email protected]");
console.log(out.safe, out.redactedText);

Using ready OGL-Mini Model

The model was trained at >96k samples to scan attacks inside your prompts (shieldlm 54k + agentic 22k + PII 15k, OWASP 2025-2026).

Note: the NPM package doesn't include a model weights. Please, download models from GitHub or Huggingface

Using at Node.js runtime

Requires onnxruntime-node package

import { HybridGuard } from "hybrid-ai-guard";

// Variant 1: use fabric to load model (recommended - load OGL-Mini model async, with fallback without ONNX-based model)
const guard = await HybridGuard.create({ modelPath: "../../models/ogl-mini/ogl-mini.onnx" });

// Variant 2: setup OGL-Mini model into scorer manually
import { createOnnxScorer } from "@ogl-mini/guard";
const scorer = await createOnnxScorer("../../models/ogl-mini/ogl-mini.onnx"); // FS path
const guard2 = new HybridGuard();
guard2.getClassifier().setOnnxScorer(scorer);

await guard.checkInput("Ignore previous instructions");

Browser (2.8MB INT8, WASM)

Requires onnxruntime-web package. For browser use quant ogl-mini.int8.onnx model (2.8MB), to avoid 250MB downloading.

// public/models/ogl-mini.int8.onnx - put file into public and use it as static content
const guard = await HybridGuard.create({ modelUrl: "/models/ogl-mini.int8.onnx" });
// or
import { createOnnxScorer } from "hybrid-ai-guard";
const scorer = await createOnnxScorer("/models/ogl-mini.int8.onnx"); // URL

If OGL-Mini model doesn't loaded, guard automatically switch to lightweight-classifier.

PII and Custom models

Using PII detector from OGL-Mini ogl-mini-pii.onnx

You can also use our PII detection model (ogl-mini-pii.onnx) in TypeScript. The module will use hybrid: regex for span + ONNX Model for reranking confidence (11 types: EMAIL, PHONE, PERSON, IP, IBAN, BANK_CARD, PASSPORT, GOV_ID, DOB, ADDRESS, SOCIAL).

Basic Usage:

// Node (2MB FP32)
import { HybridGuard } from "hybrid-ai-guard";
const guard = await HybridGuard.create({
  modelPath: "../../models/ogl-mini/ogl-mini.onnx",             // guard 250MB
  piiModelPath: "../../models/ogl-mini/ogl-mini-pii.onnx",      // PII 2MB
});

//  Or in Browser (2MB INT8)
const guardBrowser = await HybridGuard.create({
  modelUrl: "/models/ogl-mini.int8.onnx",
  piiModelUrl: "/models/ogl-mini-pii.int8.onnx",
});

// Manual initialization
import { createPiiOnnxScorer } from "hybrid-ai-guard";
const piiScorer = await createPiiOnnxScorer("../../models/ogl-mini/ogl-mini-pii.onnx");
new HybridGuard().getPii()!.setOnnxScorer(piiScorer);
// or
const pii = new PIIDetector();
pii.setOnnxScorer(await createPiiOnnxScorer("./ogl-mini-pii.onnx"));
await pii.detectAsync("Иван Петров [email protected] +7 999 123-45-67"); // reranked scores

If PII model not loaded, used only regex functions (0.5ms)

Connect custom scorer

You can use custom scorer (ONNX based), for example for fine-tuned OGL-Mini models:

import { MiniClassifier } from "hybrid-ai-guard";
const clf = new MiniClassifier({ threshold: 0.55 });
clf.setOnnxScorer(async (text) => myCustomModel.predict(text)); // () => P(attack)
const guard = new HybridGuard({ classifier: clf });

API

Basic API

• new HybridGuard(opts?) - create instance without model loading
• HybridGuard.create(opts) - create instance, with model loading
• guard.checkInput(text, withPii?) / checkInputSync(text) / checkOutput(text) / detectPii(text) - use it for security scanning
• guard.getClassifier().setOnnxScorer(fn) / guard.getHeuristics() / guard.getPii() - use it to get current modules

Pacakge Scripts

npm install
npm run build        # tsc → dist/
npm test             # vitest, 30 tests (guard, pii, performance, onnx)
npm run bench        # p50/p95 (heuristic 0.001ms, guard 0.01ms)

By-Stage project structure

• src/heuristics.ts - Stage 1. Security scanning using Heuristics modules.
• src/classifier.ts - Stage 2. Use real ML model to scan inputs.
• src/pii.ts - Stage 3. Detect PII data.
• src/onnx.ts - Used to load OGL-Mini models (ONNX based) into browser / NodeJS environment.
• src/guard.ts - General class

License

OGL-Mini library is distributed under the MIT license. You can use it however you like. I would appreciate any feedback and suggestions for improvement. Full license text can be found here

About | Installation | Get Started | More Information | Contacts