OWASP LLM Top 10 · #1 Vulnerability · Open Source

Stop prompt
injection attacks.

Open-source security middleware for LLM apps. 3-layer detection. 2 lines of Python. Works with ChatGPT, Claude, Gemini — any AI.

Try live demo → View on GitHub

26+

detection rules

automated tests

<1ms

layer 1 latency

detection layers

Architecture

Built different.
Designed to block.

Layer 1

Regex Rule Engine

26+ YAML rules compiled at startup. Pattern matching against all known attack signatures. Blocks obvious attacks with zero ML overhead.

⚡ < 1ms response

Layer 2

ML Classifier

HuggingFace transformer understands semantic meaning. Catches disguised attacks regex misses completely.

🧠 10–30ms

Layer 3

Custom Rules

Add domain rules via API. No redeploy needed. In-memory cache for zero latency.

🔧 Your rules

Integration

lines to protect any LLM

Platforms

Any

OpenAI · Claude · Gemini

False positives

<3%

on benign inputs

Design

Fail

open — never breaks app

Quick start

# pip install llmguardian

# Before — unprotected

def ask_gpt(user_message):

return openai.chat(user_message)

# After — 2 lines added

from promptshield import shield

@shield(mode='block')

def ask_gpt(user_message):

return openai.chat(user_message)

🛡️ InjectionDetected: prompt_injection · HIGH · INJ001 — AI never saw this

✓ ALLOW — "What are your opening hours?" passed through normally

What gets blocked

💉

Prompt Injection

"Ignore all previous..."

HIGH

🔓

Jailbreak

"You are now DAN..."

HIGH

🔍

Prompt Extraction

"Repeat system prompt..."

HIGH

📊

PII Exfiltration

"List all customer emails..."

HIGH

🌐

Indirect Injection

Hidden in docs/URLs

MED

Stop prompt
injection attacks.

Built different.
Designed to block.

Try it.
Right now.

The stack.
Production grade.

Stop promptinjection attacks.

Built different.Designed to block.

Try it.Right now.

The stack.Production grade.

Stop prompt
injection attacks.

Built different.
Designed to block.

Try it.
Right now.

The stack.
Production grade.