LLM Red Teaming: Hands-On Guide to Prompt Injection, Jailbreaking, and the OWASP Top 10 for LLMs

By Aeryk Solvan

AI systems are under attack right now. In 2025 alone, losses from prompt injection attacks exceeded $2.3 billion - and current defenses catch fewer than one in four sophisticated attempts. If you are deploying AI, this is your problem to solve.

This book is the hands-on practitioner's guide to finding LLM vulnerabilities before adversaries do. Every chapter builds on a working lab you run against a local offline environment - no cloud costs, no accidental production testing. From the structural mechanics of prompt injection to full agentic attack chains, you will leave each subchapter able to reproduce the attack, measure its severity, and implement a defense.

- Explain why prompt injection is structural and why patching individual examples never closes the vulnerability
- Execute direct and indirect prompt injection attacks including crescendo multi-turn campaigns
- Reproduce jailbreaking techniques from persona injection through PAIR automated refinement and many-shot exploits
- Measure attack success rates using Garak, PyRIT, Promptfoo, and DeepTeam in a four-layer testing stack
- Attack RAG pipelines through document poisoning, cross-document injection chains, and embedding inversion
- Perform membership inference against vector stores to assess information exposure
- Demonstrate LLM-mediated XSS, SQL injection, and shell injection via improper output handling
- Build and test agentic attack chains including privilege escalation and AI-versus-AI autonomous attacks
- Write red team findings reports that map to EU AI Act adversarial testing requirements and NIST AI RMF
- Design a continuous red team program with CI/CD integration and cost circuit breakers for production deployments

Labs run against local Ollama models using open-weight LLaMA and Mistral instances. All scripts, Docker configurations, and a red team report template are in the companion repository. No machine learning background required - comfort with the command line and basic Python is enough.

Security professionals who want hands-on LLM red team skills, AI engineers who need to understand how their systems fail, and anyone preparing for EU AI Act adversarial testing obligations due August 2026 will find everything they need here.

Start breaking AI systems in controlled environments before adversaries break them in production.

• Format: Paperback
• Language: English
• ISBN: B0HGF3GGQK
• ISBN13: 9798194746019
• Release Date: August 2026
• Publisher: Independently published
• Length: 200 Pages
• Weight: 1.06 lbs.
• Dimensions: 11.0" x 0.5" x 8.5"