DeepTeam
Open-source LLM red-teaming framework: penetration testing for AI agents and RAG pipelines.
About
DeepTeam by Confident AI (YC W25) is an open-source red-teaming framework for LLM systems, described as penetration testing for LLMs. It dynamically generates adversarial inputs at test time instead of relying on static datasets, simulating attacks like jailbreaking, prompt injection, encoding obfuscation, and multi-turn escalation.
It covers 40+ vulnerability categories (PII leakage, SQL injection, bias, toxicity, agentic misuse) and maps findings to OWASP LLM Top 10 and NIST AI RMF for standards-driven risk assessment. It runs locally on your machine, is built on the DeepEval evaluation framework, and also offers guardrails to prevent discovered issues in production.
What it can do
- Dynamic adversarial attack generation across 20+ attack methods
- 40+ vulnerability classes mapped to OWASP LLM Top 10 and NIST AI RMF
- Runs locally, built on the DeepEval framework
- Guardrails to block discovered vulnerabilities in production
Setup
pip install deepteam, then write a red-team run targeting your model endpoint or agent.
Prerequisites
Python and access to the LLM system you want to test (API endpoint, agent, or RAG pipeline).
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 106; Trending score 78/100 from researcher assessment.
★ 2,993 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.