# DeepTeam

> Open-source Python framework for red teaming LLM applications and AI agents by simulating adversarial attacks and detecting vulnerabilities.

DeepTeam is an open-source red teaming framework for LLM systems from Confident AI, the team behind DeepEval. It simulates attacks such as jailbreaking, prompt injection and multi-turn exploitation against AI agents, RAG pipelines and chatbots, and reports vulnerabilities like bias, PII leakage and SQL injection. It runs locally and is licensed under Apache 2.0.

## What It Is

DeepTeam works like penetration testing for LLM applications. You wrap your application in a model callback, choose vulnerabilities and attacks (or an entire safety framework), and the tool generates adversarial prompts dynamically, so no prepared dataset is needed. Each vulnerability is scored by LLM-as-a-judge metrics that run locally and produce binary pass/fail results with reasoning. The resulting risk assessment can be viewed in dataframes and saved locally as JSON.

## Vulnerabilities and Attacks

The README documents 50+ vulnerabilities grouped into data privacy, responsible AI, security, safety, business and agentic categories, plus custom vulnerabilities defined in code. The homepage advertises 120+ vulnerabilities. More than 20 attack methods cover single-turn techniques (prompt injection, roleplay, encoding obfuscation, context flooding) and multi-turn jailbreaks (Linear, Tree, Crescendo, Sequential, Bad Likert Judge).

## Frameworks and Guardrails

Built-in framework support maps categories to the relevant vulnerabilities and attacks: OWASP Top 10 for LLMs 2025, OWASP Top 10 for Agents 2026, NIST AI RMF, MITRE ATLAS, BeaverTails and Aegis. After testing, seven guardrails (toxicity, prompt injection, privacy, illegal content, hallucination, topical and cybersecurity) can classify inputs and outputs in production.

## Setup Path

Install with pip, set an OpenAI API key or configure a custom model, and run a Python script that calls red_team. Runs can also be driven from the CLI with YAML configs, and the site lists CI/CD runners such as GitHub Actions, GitLab CI and Jenkins. Results can optionally be managed in the separate Confident AI platform.

## Features
- 50+ ready-to-use vulnerabilities with LLM-as-a-judge scoring
- 20+ single-turn and multi-turn adversarial attacks
- Framework-based red teaming (OWASP, NIST, MITRE ATLAS, BeaverTails, Aegis)
- 7 production guardrails for inputs and outputs
- Custom vulnerabilities and attacks
- CLI with YAML configs and Python API
- Local risk assessment output as dataframes and JSON
- Runs locally

## Integrations
DeepEval, Confident AI, OpenAI, Claude, Gemini, Azure OpenAI, AWS Bedrock, Vertex AI, Mistral, LiteLLM, Portkey, GitHub Actions, GitLab CI, Jenkins, CircleCI, Buildkite, Azure Pipelines

## Platforms
API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
v1.0.9

## Links
- Website: https://trydeepteam.com
- Documentation: https://www.trydeepteam.com/docs/getting-started
- Repository: https://github.com/confident-ai/deepteam
- EveryDev.ai: https://www.everydev.ai/tools/deepteam