
AI Red Teaming: Test, Break, and Secure Large Language Models
The Practitioner's Playbook for Adversarial Prompting, Jailbreak Detection, Prompt Injection Attacks, and LLM Vulnerability AssessmentBy Caroline LennoxLength3h 54m
About this audiobook
Your AI application is one clever prompt away from being compromised. Do you know how to stop it?
Large language models don't break like traditional software. They can be jailbroken with a roleplay, tricked into leaking their system prompt, poisoned through the data they read, and manipulated into leaking training data or executing dangerous tool calls — and none of it looks like a normal attack.
AI Red Teaming is the hands-on playbook for security professionals, ML engineers, and developers who need to test and defend LLM systems in the real world. No hand-waving, no theory dumps — just the techniques attackers actually use, and the defenses that actually work.
Inside, you'll learn how to:
Exploit prompt injection — the "SQL injection of AI" — and detect it before attackers do
Bypass safety guardrails with jailbreak techniques like DAN, encoding attacks, and many-shot framing
Extract system prompts, training data, and PII through targeted probing
Attack multimodal models and LLM agents that use tools, APIs, and code execution
Automate red teaming at scale with Garak, PyRIT, and custom Python harnesses
Build defense-in-depth with Llama Guard, NeMo Guardrails, input validation, and output filtering
Map your findings to the NIST AI RMF and EU AI Act — and write red-team reports that hold up
Every chapter is built around working code, real tools, and a local lab you set up yourself. Whether you're securing a production LLM, building an enterprise red-team program, or breaking into AI security, this is the field manual you'll keep coming back to.
Stop hoping your AI is secure. Start proving it.
Audiobook details
GenreTechnology
Length3 hrs 54 mins
Narrated byListen with 1,000+ voices
FormateBook with Audio
LanguageEnglish
Table of contents
1Introduction
33PII Leakage Assessment and Mitigation
2CHAPTER 1: Why LLMs Break Differently Than Everything Else
34CHAPTER 7: Adversarial Inputs for Multimodal Models
3The New Attack Surface
35Vision-Language Models: New Attack Surface
4Classical Vuln vs LLM Vuln: A Side-by-Side
36Adversarial Patches and Pixel-Level Attacks
5The Four Failure Modes: Hallucination, Injection, Extraction, Misuse
37Text-in-Image Prompt Injection (Visual Injection)
Show all chaptersShow less
6Threat Modeling an LLM Application
38Cross-Modal Consistency Testing
7CHAPTER 2: Red Teaming Fundamentals and Lab Setup
39CHAPTER 8: LLM Agent and Tool-Use Attacks
8Lab Architecture: Local vs API-Based Testing
40How LLM Agents Use Tools and Why That's Dangerous
9Installing Ollama and Running Llama 3 Locally
41Tool Output Injection: Hijacking the Agent's Next Action
10Garak and PyRIT: Automated LLM Vulnerability Scanners
42SSRF, Command Injection, and Path Traversal via Agent Tools
11Rules of Engagement and Responsible Disclosure
43Multi-Agent Attack Chains: Privilege Escalation Across Agents
12CHAPTER 3: Prompt Injection: The SQL Injection of AI
44CHAPTER 9: Automated Red Teaming at Scale
13Direct Injection: Overriding the System Prompt
45Garak Deep Dive: Probes, Generators, and Detectors
14Indirect Injection: Poisoning External Data Sources
46LLM-as-Attacker: Using One LLM to Red-Team Another
15Role Confusion and Persona Hijacking Techniques
47Building a Custom Attack Harness with Python
16Detection: Signature-Based and Semantic Classifiers
48Triage Framework: Severity Scoring for LLM Vulnerabilities
17CHAPTER 4: Jailbreaking: Bypassing Safety Guardrails
49CHAPTER 10: Defenses: Guardrails, Input Validation, and Output Filtering
18Taxonomy of Jailbreaks: Persona, Encoding, Framing, Many-Shot
50Defense-in-Depth Architecture for LLM Applications
19DAN, AIM, and Roleplay-Based Bypasses
51Input Validation: Semantic Classifiers and Pattern Matching
20Encoding Attacks: Base64, ROT13, Token Smuggling
52Llama Guard and NeMo Guardrails in Practice
21Code: Base64 and Token-Smuggling Payload Generator
53Output Filtering: PII Redaction, Toxicity, and Hallucination Detection
22Code: Jailbreak Benchmark Harness
54CHAPTER 11: Compliance: NIST AI RMF and EU AI Act
23Why Safety Training Fails: Competing Objectives
55NIST AI RMF: Govern, Map, Measure, Manage
24CHAPTER 5: System Prompt Extraction and Leakage
56EU AI Act: Prohibited Practices and High-Risk Obligations
25How System Prompts Work and Why They Leak
57Mapping Red-Team Findings to Compliance Controls
26Extraction Techniques: Repeat, Summarize, Translate
58Writing the Red-Team Report: Structure and Evidence Standards
27Detecting System Prompt Presence from Response Patterns
59CHAPTER 12: Building an Enterprise AI Red-Team Program
28Hardening: What Works and What Doesn't
60The AI Red-Team Lifecycle: Scope → Attack → Report → Verify
29CHAPTER 6: Model Inversion and Training Data Extraction
61Integrating Red-Teaming into CI/CD with Garak and GitHub Actions
30Memorization in LLMs: Why Models Remember Verbatim Text
62Building the Team: Skills, Tools, and Hiring Profile
31Prefix Attacks for Training Data Extraction
63The Road Ahead: Agentic AI and the Next Generation of Attacks
32Membership Inference: Did This Text Train the Model?
64About the Author