LLM Red Team Testing Playbook: A Reproducible Adversarial Detection Toolkit Based on OWASP and MITRE Frameworks

An open-source toolkit for AI security researchers and red team engineers, providing reproducible adversarial detection mapped to OWASP LLM Top 10 Risks 2025 and MITRE ATLAS technical framework.

• Original Author/Maintainer: Leonardo Jaguaribe (OSCP-certified cybersecurity instructor)
• Source Platform: GitHub
• Original Title: llm-redteam-playbook
• Original Link: https://github.com/leojaguaribe/llm-redteam-playbook
• Publication Date: 2026-05-24

---

The field of red team testing for Large Language Models (LLMs) is evolving rapidly, but most public resources remain at the level of marketing slides or one-off Twitter threads. This situation makes it difficult for security practitioners to obtain reproducible and quantifiable test results, preventing them from truly assessing the security risks of LLMs in production environments.

Leonardo Jaguaribe's open-source llm-redteam-playbook project was created to address this issue. It provides a small, runnable, and opinionated detection toolkit that allows security practitioners to prove "this model has a vulnerability in LLM01 (Prompt Injection) today" via the command line instead of PPT presentations.

---

This playbook targets three core user groups:

AI Security Researchers: Need a systematic framework to study the adversarial robustness of LLMs

Red Team Engineers: Need reproducible testing tools to evaluate enterprise-deployed LLM systems

Machine Learning Security Practitioners: Professionals who need to align security testing with industry-standard frameworks

The core design philosophy of the project is "executable evidence, not slides"—each detection can be reproduced within two minutes after installation, providing specific and verifiable security findings.

---

The project fully covers the OWASP-released Top 10 Security Risks for LLM Applications 2025, with each risk category corresponding to a dedicated detection module:

Current Status: v0.0.1 basic version implemented

Prompt injection is the most fundamental and dangerous attack vector in LLM security. Attackers attempt to override system prompts or manipulate model behavior through carefully crafted inputs. This detection module tests the model's ability to identify boundaries for malicious inputs.

Current Status: Planned

Tests whether the model inadvertently leaks sensitive information from training data, such as personally identifiable information (PII), trade secrets, or other confidential content.

Current Status: Planned

Evaluates the security risks posed by the model's reliance on third-party components, plugins, or external data sources.