AI Guard Python SDK

Runtime security for LLM applications. Detects prompt injection, unsafe output, and other AI security threats in real-time.

Quickstart

pip install scan5-ai-guard

from ai_guard import AIGuardSDK, CapturedRequest, CapturedResponse, ModelMetadata

sdk = AIGuardSDK({"enabled": True})

# Scan a request/response pair
findings = await sdk.scan_request_response(
    CapturedRequest(
        id="req-123",
        url="openai:chat.completions.create",
        body={"messages": [{"role": "user", "content": "Hello!"}]},
    ),
    CapturedResponse(
        status_code=200,
        body={"choices": [{"message": {"content": "Hi there!"}}]},
        duration_ms=150,
    ),
    ModelMetadata(
        name="gpt-4",
        provider="openai",
    ),
)

for finding in findings:
    print(f"[{finding.severity}] {finding.message} (confidence={finding.confidence:.2f})")

Provider Wrappers

Automatically scan OpenAI and Anthropic API calls:

import openai
from ai_guard import AIGuardSDK, wrap_openai_client

sdk = AIGuardSDK()
client = openai.OpenAI()
wrapped = wrap_openai_client(client, sdk)

# All chat completions are scanned automatically
response = wrapped.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "Tell me a joke"}],
)

import anthropic
from ai_guard import AIGuardSDK, wrap_anthropic_client

sdk = AIGuardSDK()
client = anthropic.Anthropic()
wrapped = wrap_anthropic_client(client, sdk)

response = wrapped.messages.create(
    model="claude-3-opus-20240229",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)

Zero-Config Philosophy

The SDK works out of the box with sensible defaults. Configure via environment variables:

Configuration

sdk = AIGuardSDK({
    "enabled": True,
    "evaluation_timeout_ms": 100,
    "sample_rate_percent": 100,
    "capture_payloads": False,
    "environment": "production",
    "classifier_threshold": 0.85,
    "classifier_enabled": True,
    "on_finding": lambda finding: print(f"Finding: {finding}"),
    "on_error": lambda error: print(f"Error: {error}"),
})

Dashboard

View findings and manage configuration at https://scan5.vercel.app. Findings are transmitted to the dashboard when AI_GUARD_PLATFORM_URL and AI_GUARD_RUNTIME_INGEST_SIGNING_KEY are configured.

API

AIGuardSDK

Development

# Install in editable mode
pip install -e .

# Install with provider extras
pip install -e ".[openai,anthropic]"

# Build distribution packages
pip install build
python -m build

License

MIT