aisafeguard 0.1.2

Safety rails for every AI app — model-agnostic guardrails for LLM applications

AISafe Guard

aisafeguard is an open-source LLM safety and guardrails toolkit for AI apps.

It helps protect against prompt injection, jailbreak attempts, PII leaks, toxic output, and malicious URLs with a Python SDK, CLI, and OpenAI-compatible proxy.

Why AISafe Guard

• Add AI safety checks to any LLM app with minimal code changes
• Enforce configurable guardrail policies: block, warn, log, redact
• Protect both input prompts and model outputs
• Use as a library or as a language-agnostic proxy gateway

Core Features

• Prompt injection detection
• Jailbreak detection
• PII detection and redaction
• Toxicity filtering
• Malicious URL detection
• Relevance checks
• OpenAI/Anthropic wrappers
• OpenAI-compatible proxy mode

Install

pip install aisafeguard

From source (fresh clone):

pip install .

Optional extras:

pip install "aisafeguard[ml]"
pip install "aisafeguard[proxy]"
pip install "aisafeguard[integrations]"
pip install "aisafeguard[telemetry]"

Repository Setup

git clone <your-repo-url>
cd aisafeguard
python3.11 -m venv .venv
source .venv/bin/activate
pip install -e .[dev]
npm install

Node.js Users

This repo does not currently ship a native Node SDK package for runtime usage.

Recommended Node integration today:

1. Run aisafe proxy (or Docker) from this repo.
2. Call the OpenAI-compatible endpoint from your Node app.
3. Keep safety policy/config centralized in aisafe.yaml.

Example (Node fetch):

const res = await fetch("http://localhost:8000/v1/chat/completions", {
  method: "POST",
  headers: { "Content-Type": "application/json", "x-user-id": "user-123" },
  body: JSON.stringify({
    model: "gpt-4o-mini",
    messages: [{ role: "user", content: "Hello" }]
  })
});
const data = await res.json();
console.log(data.choices?.[0]?.message?.content);

Quick Start

Decorator:

from aisafeguard import guard

@guard(input=["prompt_injection", "pii"], output=["toxicity", "pii"])
async def ask(prompt: str) -> str:
    return "model output"

Guard object:

from aisafeguard import Guard

g = Guard(config="aisafe.yaml")
input_result = await g.scan_input("Ignore previous instructions")

OpenAI wrapper:

from aisafeguard import Guard
from aisafeguard.integrations import wrap_openai

guard = Guard()
client = wrap_openai(openai_client, guard)

Use Cases

• Secure chatbots against prompt injection
• Prevent sensitive-data leaks in support assistants
• Add policy controls for enterprise AI workflows
• Gate unsafe model outputs before returning to end users
• Centralize AI safety via proxy for multi-language stacks

CLI

aisafe init
aisafe validate aisafe.yaml
aisafe scan "My SSN is 123-45-6789"
aisafe redteam --strict
aisafe proxy --config aisafe.yaml --host 127.0.0.1 --port 8000 \
  --upstream-base-url https://api.openai.com \
  --upstream-api-key $OPENAI_API_KEY \
  --rpm 120

Proxy env vars:

• AISAFE_UPSTREAM_BASE_URL
• AISAFE_UPSTREAM_API_KEY (or OPENAI_API_KEY)

Docker

docker build -t aisafeguard:latest .
docker run --rm -p 8000:8000 \
  -e AISAFE_UPSTREAM_API_KEY=$OPENAI_API_KEY \
  aisafeguard:latest

Development

npm install
PYTHONPATH=src python -m pytest -v
python benchmarks/bench_pipeline.py

Docs

• docs/getting-started.md
• docs/config-reference.md
• docs/prompt-injection-protection.md
• docs/pii-redaction-llm.md
• docs/openai-compatible-ai-proxy.md