Welcome to opensentry Discussions!

[sharma-open-source]

🚀 Introducing OpenSentry – An Open-Source Security Layer for LLM Applications

Hi everyone! 👋

I've been working on an open-source project called OpenSentry, and I'd love to get feedback from the community.

The idea came from a simple observation.

For traditional software, we've spent years adopting security best practices:

• Input validation
• SQL injection prevention
• XSS protection
• Authentication & authorization

But with AI applications, many architectures still look like this:

User
  │
  ▼
LLM

Raw user input is forwarded directly to the model without any validation or inspection.

As prompt injection attacks, jailbreaks, hidden instructions, and RAG-based attacks become more common, I felt there was a missing security layer.

That led me to build OpenSentry.

What is OpenSentry?

OpenSentry is an open-source TypeScript library that sits between users and an LLM.

Its goal is to evaluate prompts before they reach the model.

The architecture is intentionally layered:

User
   │
   ▼
OpenSentry
   ├── Tier 0 – Fast heuristics
   ├── Tier 1 – Local ONNX guard (optional)
   └── Tier 2 – Remote LLM judge (optional)
   │
   ▼
OpenAI / Claude / Gemini / Ollama / Any OpenAI-compatible API

The philosophy is to keep the common path extremely fast while allowing deeper semantic analysis only when it's actually needed.

Current Features

• Prompt injection detection
• Instruction override detection
• Unicode normalization
• Encoded payload detection
• Configurable multi-tier pipeline
• Works with any LLM provider
• Edge-runtime compatible
• Node.js, Bun, Deno, Browser support

Benchmarks

Current benchmark highlights:

• Tier 0 heuristics: ~0.02 ms p50
• Tier 1 local ONNX guard: ~6–8 ms p50
• Tier 2 remote LLM judge (optional): ~467 ms p50

The goal isn't to send every prompt to another LLM. Instead, only ambiguous or high-risk prompts are escalated to the remote tier.

Why I'm Posting

I'm not looking for stars as much as I'm looking for honest technical feedback.

Some questions I'd love opinions on:

• Does this layered architecture make sense for production AI systems?
• Are there prompt injection datasets you'd recommend adding?
• Which local moderation models have worked well in your experience?
• Are there security scenarios that OpenSentry should support but currently doesn't?

Constructive criticism is more than welcome.

GitHub Repository:
https://github.com/sharma-open-source/opensentry

Thanks for taking the time to read this—I appreciate any feedback, suggestions, or ideas for improving the project.