llmgateways 0.1.0

Protect OpenAI and Anthropic API calls from prompt injection, jailbreaks, and data-extraction attacks.

llmgateways

Python SDK for LLM Gateways — protect OpenAI and Anthropic API calls from prompt injection, jailbreaks, and data-extraction attacks.

Installation

pip install llmgateways            # core only
pip install "llmgateways[openai]"  # + OpenAI
pip install "llmgateways[anthropic]"  # + Anthropic
pip install "llmgateways[all]"     # + both

Quick start

OpenAI

from llmgateways import wrap, PromptBlockedError
from openai import OpenAI

client = wrap(OpenAI(), api_key="lgk_...")

try:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": "Hello!"}],
    )
    print(response.choices[0].message.content)
except PromptBlockedError as e:
    print(f"Blocked! Threats: {e.result.threats}")
    print(f"Risk score: {e.result.risk_score:.2f}")

Anthropic

from llmgateways import wrap, PromptBlockedError
from anthropic import Anthropic

client = wrap(Anthropic(), api_key="lgk_...")

try:
    response = client.messages.create(
        model="claude-3-5-sonnet-20241022",
        system="You are a helpful assistant.",
        messages=[{"role": "user", "content": "Hello!"}],
        max_tokens=1024,
    )
    print(response.content[0].text)
except PromptBlockedError as e:
    print(f"Blocked! Threats: {e.result.threats}")

Async

Both OpenAI and Anthropic async clients are supported:

from llmgateways import wrap, PromptBlockedError
from openai import AsyncOpenAI

client = wrap(AsyncOpenAI(), api_key="lgk_...")

async def main():
    try:
        response = await client.chat.completions.create_async(
            model="gpt-4o",
            messages=[{"role": "user", "content": "Hello!"}],
        )
    except PromptBlockedError as e:
        print("Blocked:", e.result.threats)

How it works

Every call to chat.completions.create or messages.create is intercepted:

1. The prompt is sent to the LLM Gateways detection engine
2. L1 — pattern matching (instant, <1 ms)
3. L2 — semantic similarity via MiniLM embedding model
4. L3 — LLM judge (DeepSeek) for ambiguous cases
5. If blocked → PromptBlockedError is raised before the model is called
6. If allowed → the original call proceeds unchanged

API reference

wrap(client, *, api_key, base_url="", timeout=10.0)

Returns a protected proxy with the same interface as the original client.

Parameter	Type	Description
client	OpenAI | Anthropic	The LLM client to protect
api_key	str	Your lgk_... key from the dashboard
base_url	str	Override for self-hosted deployments
timeout	float	Gateway request timeout in seconds (default: 10)

PromptBlockedError

except PromptBlockedError as e:
    e.result.risk_score   # float 0.0–1.0
    e.result.action       # "block"
    e.result.threats      # list[str], e.g. ["jailbreak", "injection"]
    e.result.layer_used   # int (1, 2, or 3)
    e.result.reasoning    # str | None (populated by L3 LLM judge)
    e.result.latency_ms   # int

LLMGatewaysClient

Use directly if you want to call the scan API without wrapping a client:

from llmgateways import LLMGatewaysClient

gw = LLMGatewaysClient(api_key="lgk_...")
result = gw.scan("Hello!", system_prompt="You are helpful", model="gpt-4o")
print(result.action)  # "allow" or "block"

# Async
result = await gw.scan_async("Hello!")

Get an API key

Sign up at llmgateways.com → Dashboard → API Keys → Create key.

License

MIT