app.py

import gradio as gr
from transformers import pipeline
import re

# Load open-source LLM
llm = pipeline("text-generation", model="gpt2")

# Define guardrails
blocked_keywords = ["kill", "bomb", "hack", "suicide", "drugs"]
blocked_patterns = [r"\bhow to (make|build|create) (a )?(bomb|weapon)\b", r"\b(hack|bypass) (password|system)\b"]

# Violation log
violation_log = []

def guardrails_check(prompt):
    for keyword in blocked_keywords:
        if keyword in prompt.lower():
            violation_log.append(f"Keyword blocked: {keyword}")
            return False
    for pattern in blocked_patterns:
        if re.search(pattern, prompt.lower()):
            violation_log.append(f"Pattern blocked: {pattern}")
            return False
    return True

def generate_response(prompt):
    if not guardrails_check(prompt):
        return "🚫 Prompt blocked by guardrails."
    output = llm(prompt, max_length=100, do_sample=True)[0]["generated_text"]
    return output

def show_log():
    return "\n".join(violation_log[-5:]) if violation_log else "No violations yet."

with gr.Blocks() as demo:
    gr.Markdown("## 🛡️ LLM Guardrails Demo")
    with gr.Row():
        prompt = gr.Textbox(label="Enter your prompt")
        output = gr.Textbox(label="LLM Response")
    with gr.Row():
        log = gr.Textbox(label="Violation Log", interactive=False)
    submit = gr.Button("Generate")
    submit.click(fn=generate_response, inputs=prompt, outputs=output)
    submit.click(fn=show_log, inputs=None, outputs=log)

demo.launch()