Deploying Large Language Models (LLMs) without rigorous security testing is a massive risk. From PII leakage to prompt injection, vulnerabilities can compromise your entire application before it even reaches production.
What is DeepTeam?
Think of DeepTeam as penetration testing specifically designed for the era of AI. It is an open-source red teaming framework that allows you to simulate complex attacks against your LLM systems, including RAG pipelines and chatbots. You can explore the DeepTeam repository on GitHub to get started with this powerful tool built on top of DeepEval.
Uncovering Vulnerabilities and Attacks
To secure an AI, you must first understand how it can be broken. If you are wondering why traditional software testing fails for AI, the answer is that LLMs require probabilistic rather than deterministic security checks.
Comprehensive Vulnerability Probes
DeepTeam provides over 50 ready-to-use vulnerabilities, each accompanied by clear explanations and LLM-as-a-Judge metrics. These probes run locally on your machine to provide binary pass/fail scores with reasoning across several critical domains:
- Data Privacy
- Responsible AI
- Security (including SQL injection)
- Safety and Bias
- Business Logic and Agentic workflows
Advanced Adversarial Attack Methods
Simple queries aren’t enough to catch sophisticated threats. DeepTeam utilizes over 20 research-backed adversarial attack methods, ranging from single-turn prompts to complex multi-turn conversational exploitation. These attacks use state-of-the-art techniques like encoding-based obfuscation and jailbreaking to test the limits of your model’s guardrails.
From Testing to Production Guardrails
Once you have identified weaknesses, you need a way to prevent them in real-time. DeepTeam offers seven production-ready guardrails designed for fast binary classification of LLM inputs and outputs. These tools help ensure your application adheres to established AI safety frameworks by mapping categories directly to specific vulnerabilities.
For those looking at the broader landscape, DeepTeam is a vital component in an ecosystem that includes tools like Strix for autonomous AI agent penetration testing. By integrating these security layers, you can move from prototype to a hardened production environment with confidence.
Implementation and Quick Start
Getting started is straightforward. First, install the package via pip:
pip install -U deepteam
You can then run a basic red teaming assessment with just a few lines of Python code. Here is how you can test for bias and prompt injection in your own model callback:
from deepteam import red_team
from deepteam.vulnerabilities import Bias
from deepteam.attacks.single_turn import PromptInjection
async def model_callback(input: str) -> str:
# Replace this with your LLM application
return f"I'm sorry but I can't answer this: {input}"
risk_assessment = red_team(
model_callback=model_callback,
vulnerabilities=[Bias(types=["race"])],
attacks=[PromptInjection()]
)
Start testing your LLM today to ensure it is secure, safe, and ready for production.
