Resk-LLM으로 Python LLM 파이프라인 보호 — 한 번에 11개의 위협 탐지기 설치

작성자

카테고리:

← 피드로
DEV Community · RESK · 2026-07-16 개발(SW)

RESK

Links

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

Prompt injection and jailbreak attacks are the most common LLM security threats today. Most defenses are either too slow for real-time use or locked behind proprietary APIs.

Resk-LLM is an open source Python toolkit that brings 11 threat detectors into a single FastAPI middleware. Install it, decorate your endpoint, and get instant protection against prompt injection, jailbreak, PII leakage, code injection, exfiltration, and more.

Installation

pip install resk-llm

Enter fullscreen mode Exit fullscreen mode

FastAPI Middleware Example

from fastapi import FastAPI
from resk_llm import RESKSecurityMiddleware

app = FastAPI()

# Attach threat detection to your LLM route
app.add_middleware(
    RESKSecurityMiddleware,
    detectors=["injection", "jailbreak", "pii", "exfiltration"],
    mode="block"  # or "log" for monitoring
)

@app.post("/chat")
async def chat(message: dict):
    # If we reach here, input passed all detectors
    response = await call_your_llm(message["content"])
    return {"response": response}

Enter fullscreen mode Exit fullscreen mode

Available Detectors

Detector What It Catches Prompt Injection DAN, ignore-prior-instructions, role-play escapes Jailbreak Hypothetical traps, obfuscated instructions PII Leakage Emails, phone numbers, SSNs, credit cards Exfiltration Prompt stealing, data dumping Toxic Content Hate speech, harassment, profanity Code Injection SQL, shell, Python eval attempts URL Phishing Malicious URL patterns And 4 more Hidden content, base64 encoding, etc.

Why Dual-Layer Defense?

Resk-LLM works on the input side, catching threats before they reach the model. For even stronger protection, pair it with resk-logits — a GPU-accelerated logits processor that blocks dangerous tokens at generation time. Combined, you get input filtering AND output sanitation in one pipeline.

Production Ready

  • MIT licensed, open source
  • Python 3.13+ and PyTorch 2.0+
  • Zero external API dependencies
  • Configurable per-route policies
  • Log mode for evaluation, block mode for production
pip install resk-llm

Enter fullscreen mode Exit fullscreen mode

Try it on your next project. Open an issue if a detector misses something — this is community-driven security, and feedback makes it better.

원문에서 계속 ↗

코멘트

답글 남기기

이메일 주소는 공개되지 않습니다. 필수 필드는 *로 표시됩니다