Secure Your Python LLM Pipeline with Resk-LLM: 11 Threat Detectors in One Middleware

#python #security #llm #opensource

Links

GitHub: https://github.com/Resk-Security/Resk-LLM
PyPI: https://pypi.org/project/resk-llm
Website: https://resk.fr

If you deploy an LLM in production you need layered security. System prompts help but they are not enough. Jailbreaks, prompt injections and exfiltration attempts can bypass instruction-based filters entirely.

Resk-LLM is an open source Python security toolkit that detects 11 categories of threats and integrates as FastAPI middleware. Lets see how easy it is to add.

Installation

pip install resk-llm

Quick Start

Add the security middleware to any FastAPI app:

from fastapi import FastAPI
from resk_llm import SecurityMiddleware

app = FastAPI()

# Enable all 11 detectors with default settings
app.add_middleware(SecurityMiddleware)

@app.post("/chat")
async def chat(prompt: str):
    # Your LLM call here
    # SecurityMiddleware handles detection automatically
    return {"response": await call_llm(prompt)}

What gets detected:

Prompt injection and jailbreak attempts
PII and sensitive data leaks
Code exfiltration and system prompt extraction
Token smuggling and adversarial suffix attacks
And 7 more categories covering the OWASP LLM Top 10

Threat Response

Each detection can be configured to:

Block: Reject the request entirely
Flag: Log the attempt and let it pass
Replace: Sanitise the offending content

resk-logits Integration

For token-level blocking pair Resk-LLM with resk-logits. Dangerous tokens are shadow-banned at the logits layer via GPU accelerated Aho-Corasick matching. The model never even generates the first token of a forbidden phrase.