CalvinBackup/ai-llm-red-team-handbook
1
0
Fork 0
mirror of https://github.com/Shiva108/ai-llm-red-team-handbook.git synced 2026-02-12 14:42:46 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
3dfd102404ecd76dcbd8701cf8eb0b57429f4c8a
ai-llm-red-team-handbook/docs/Chapter_46_Conclusion_and_Next_Steps.md

47 KiB
Raw Blame History

Chapter 46: Conclusion and Next Steps

You've reached the end of the AI LLM Red Team Handbook. But in security, there is no end—only the next model update.

46.1 The Journey So Far

Over 46 chapters, you've mastered the complete spectrum of AI security—from foundational ethics to cutting-edge algorithmic attacks. This handbook represents the culmination of field knowledge, research, and operational experience across the entire AI red teaming discipline.

The Handbook Map

graph TD
    Part1[Part I: Professional Foundations<br/>Ch 1-4] --> Part2[Part II: Project Preparation<br/>Ch 5-8]
    Part2 --> Part3[Part III: Technical Fundamentals<br/>Ch 9-11]
    Part3 --> Part4[Part IV: Pipeline Security<br/>Ch 12-13]
    Part4 --> Part5[Part V: Attacks & Techniques<br/>Ch 14-24]
    Part5 --> Part6[Part VI: Defense & Mitigation<br/>Ch 25-30]
    Part6 --> Part7[Part VII: Advanced Operations<br/>Ch 31-39]
    Part7 --> Part8[Part VIII: Strategic Topics<br/>Ch 40-46]

Handbook Structure Map

What You've Mastered

Part I: Professional Foundations (Ch 1-4)

  • The fundamentals of AI red teaming and why it's critical
  • Ethical frameworks, legal compliance (CFAA, GDPR), and stakeholder communication
  • The red teamer's mindset: curiosity, creativity, and systematic thinking
  • SOW, Rules of Engagement, and client onboarding procedures

Part II: Project Preparation (Ch 5-8)

  • Threat modeling methodologies specific to AI systems
  • Scoping engagements: in-scope vs out-of-scope for AI components
  • Building isolated red team labs with proper safety controls
  • Chain of custody and evidence handling for non-deterministic systems

Part III: Technical Fundamentals (Ch 9-11)

  • LLM architectures: transformers, attention mechanisms, tokenization
  • Context windows, generation parameters, and model behavior
  • Plugin architectures and API integration points

Part IV: Pipeline Security (Ch 12-13)

  • RAG system vulnerabilities and retrieval poisoning
  • Supply chain attacks: from training data to model distribution
  • Vector database security and embedding space manipulation

Part V: Attacks & Techniques (Ch 14-24)

  • Prompt injection (direct, indirect, multi-turn)
  • Data leakage and training data extraction
  • Jailbreaking and content filter bypass
  • Plugin exploitation and API abuse
  • Evasion techniques against detection systems
  • Data poisoning and model manipulation
  • Model theft and extraction attacks
  • Denial of service (token exhaustion, sponge attacks)
  • Multimodal attacks (vision, audio, cross-modal)
  • Social engineering via AI agents

Part VI: Defense & Mitigation (Ch 25-30)

  • Adversarial machine learning and robustness
  • Supply chain defense and model provenance
  • Federated learning security
  • Privacy-preserving techniques
  • Model inversion defenses
  • Backdoor detection and mitigation

Part VII: Advanced Operations (Ch 31-39)

  • OSINT and reconnaissance for AI systems
  • Attack automation frameworks (TAP, GCG)
  • Automated red team pipelines
  • Defense evasion and anti-detection
  • Post-exploitation in AI environments
  • Professional reporting and CVSS for AI
  • Remediation strategies and patch validation
  • Continuous red teaming programs
  • Bug bounty hunting for AI vulnerabilities

Part VIII: Strategic Topics (Ch 40-46)

  • Compliance frameworks (EU AI Act, ISO 42001, NIST AI RMF)
  • Industry best practices and defense architectures
  • Real-world case studies and failure analysis
  • Future of AI red teaming (algorithmic attacks, formal verification)
  • Emerging threats (Shadow AI, deepfakes, critical infrastructure)
  • Building organizational red team programs
  • This conclusion: your path forward

Your New Capabilities

You can now:

  1. Assess any AI system for security vulnerabilities across the full attack surface
  2. Execute sophisticated attacks from prompt injection to model extraction
  3. Automate red team operations using Python, Garak, PyRIT, and custom tools
  4. Report findings in compliance-friendly formats (ISO 42001, NIST AI RMF, EU AI Act)
  5. Design defense-in-depth architectures with input/output guardrails
  6. Build continuous red team programs with metrics and automation
  7. Navigate the legal and ethical landscape of AI security testing
  8. Hunt high-value bugs in bounty programs with proven methodologies

46.2 Staying Current in a Rapidly Evolving Field

AI security changes faster than any other domain. New models drop monthly, new attacks weekly, new defenses daily. Here's your strategy to stay ahead.

Daily Engagement (5-10 minutes)

Twitter/X Accounts to Follow

  • @AIatMeta - Meta AI research announcements
  • @OpenAI - Model releases and safety research
  • @AnthropicAI - Constitutional AI and alignment research
  • @GoogleDeepMind - Cutting-edge ML security
  • @simonw - AI engineering and security insights
  • @goodside - Creative jailbreaks and prompt research
  • @mmitchell_ai - AI ethics and responsible development
  • @random_walker - Privacy and AI policy

Security News Aggregators

Weekly Deep Dives (30-60 minutes)

ArXiv Preprints

Security Blogs

Podcasts

  • "Latent Space" - AI engineering and security
  • "Practical AI" - Applied ML and deployment challenges
  • "Security Now" (AI segments) - Broader security context

Monthly Learning (2-4 hours)

Conference Recordings

  • DEF CON AI Village (YouTube)
  • Black Hat AI Security Summit
  • NeurIPS Red Teaming Competition
  • ICLR Workshop on Secure and Trustworthy ML
  • USENIX Security (ML Security Track)

Technical Deep Reads

  • New model "System Cards" (GPT-4, Claude 3, Gemini) - Read the red team appendix
  • OWASP Top 10 for LLM Applications updates
  • MITRE ATLAS framework additions
  • New CVEs in ML frameworks (PyTorch, TensorFlow, HuggingFace)

Hands-On Practice

Certifications and Formal Training

Relevant Certifications

  • OSCP (Offensive Security Certified Professional) - Foundational pentesting + AI augmentation
  • GIAC (GXPN, GWAPT) - Advanced exploitation skills applicable to AI
  • Cloud (AWS/Azure/GCP ML Specialty) - Understanding AI deployment infrastructure
  • Custom: Several organizations (SANS, etc.) are developing AI-specific security certs - watch for these in 2025-2026

University Programs

  • Stanford CS 329S (ML Systems Security) - Free online materials
  • UC Berkeley CS 294 (AI Security) - Public lectures
  • Carnegie Mellon INI course on Adversarial ML

Community Engagement

Discord Servers

  • AI Safety - General AI alignment and security
  • HuggingFace - ML engineering and model security
  • MLSecOps - ML security operations

Slack Communities

  • OWASP AI Security & Privacy Slack
  • AI Village (DEF CON)

GitHub Monitoring

  • Watch repositories: garak-ai/garak, Azure/PyRIT, NVIDIA/NeMo-Guardrails
  • Follow security researchers publishing LLM attack tools

46.3 Contributing to the Handbook

This handbook is open source and living. Your contributions make it better.

How to Contribute

Repository: github.com/Shiva108/ai-llm-red-team-handbook

Contribution Types

  1. Bug Fixes & Typos

    • Fork the repo
    • Fix the issue
    • Submit PR with clear description

  2. New Attack Techniques

    • Follow Chapter Template format (docs/templates/Chapter_Template.md)
    • Include working code examples
    • Provide 3+ research citations
    • Add to appropriate Part in README

  3. Tool Integrations

    • Add to Appendix C with installation instructions
    • Provide quick-start example
    • Link to official documentation

  4. Case Studies

    • Use real-world incidents (anonymized if needed)
    • Include timeline, technical details, lessons learned
    • Can be added to Chapter 42 or as standalone in case_studies/

  5. Translations

    • Contact maintainers first to coordinate
    • Maintain technical accuracy
    • Keep code examples in English with translated comments

Quality Standards

  • All code must be tested and working
  • Citations required for claims
  • Follow existing markdown style
  • Pass markdownlint checks
  • No plagiarism - original work or properly attributed

46.4 Career Pathways in AI Security

AI security is one of the highest-growth career tracks in cybersecurity. Here's your roadmap.

Job Titles and Roles

Entry Level ($80k-$120k)

  • Junior AI Security Engineer
  • ML Security Analyst
  • AI Red Team Associate
  • LLM Security Researcher (Junior)

Mid-Level ($120k-$180k)

  • AI Security Engineer
  • Senior ML Security Researcher
  • AI Red Team Lead
  • Prompt Security Specialist
  • AI Safety Engineer

Senior/Staff ($180k-$300k+)

  • Principal AI Security Architect
  • Director of AI Red Teaming
  • Head of AI Safety
  • AI Security Consultant (Independent)
  • Bug Bounty Professional (AI Specialist)

FAANG Levels

  • L4/L5 (Mid): $150k-$250k total comp
  • L6/L7 (Senior/Staff): $300k-$500k+ total comp
  • L8+ (Principal): $500k-$1M+ total comp

AI Security Career Ladder

Building Your Portfolio

Must-Haves

  1. Public GitHub Repository

    • Custom AI security tools (scanner, fuzzer, analyzer)
    • Automated red team scripts
    • Contributions to Garak, PyRIT, or similar projects
    • Well-documented, production-quality code

  2. Technical Writeups

    • Medium/personal blog with deep technical analysis
    • 3-5 detailed posts on:
      • Novel attack technique you discovered
      • Tool you built and how it works
      • Case study of interesting vulnerability
      • Defense architecture you implemented
    • Clear writing, code snippets, diagrams

  3. Bounties or CVEs

    • Even 1-2 valid reports show real-world skill
    • Document methodology in writeups (after disclosure period)
    • OpenAI, Google, Microsoft most prestigious

  4. Conference Talks or CTF Wins

    • DEF CON AI Village lightning talk
    • Local BSides presentation
    • Top 10 finish in Gandalf or Crucible CTF
    • HackThePrompt.com leaderboard mention

Nice-to-Haves

  • Research paper (even if just arXiv preprint)
  • YouTube channel with technical tutorials
  • Open source tool with 100+ GitHub stars
  • Active participation in OWASP AI Security group

Interview Preparation

Common Technical Questions

  1. Architecture: "Explain how a RAG system works and identify 3 attack vectors."
  2. Hands-On: "Here's a system prompt. Show me 3 ways to leak it."
  3. Tooling: "Walk me through how you'd use Burp Suite to test an LLM API."
  4. Compliance: "How would you map a prompt injection finding to ISO 42001 controls?"
  5. Defense: "Design a defense-in-depth architecture for a customer support chatbot."
  6. Scenario: "You found a training data extraction vulnerability. Walk me through responsible disclosure."

Behavioral Questions

  • "Tell me about a time you found a critical vulnerability. How did you report it?"
  • "How do you stay current in AI security given the pace of change?"
  • "Describe a disagreement with a client/team about security priorities. How did you resolve it?"

Reverse Interview Questions (Ask Them)

  • "What does your AI red team program look like today? What are you building toward?"
  • "How much of my time would be manual testing vs tool development vs research?"
  • "What's your approach to responsible disclosure when you find issues in third-party models?"
  • "How do you balance 'move fast' engineering culture with security rigor?"

Salary Negotiation Tips

  • Use levels.fyi to research compensation bands
  • AI security roles often get 10-20% premium over general AppSec
  • Remote roles from FAANG/top startups pay the same globally
  • Equity/RSUs matter - $50k base difference meaningless if equity 3x larger
  • Negotiate based on competing offers, not "what you need"

46.5 The 30/60/90 Day Action Plan

From handbook reader to practicing AI red teamer in 90 days.

gantt
    title Your 90-Day AI Security Journey
    dateFormat  YYYY-MM-DD
    section Foundation
    Setup Lab Environment           :a1, 2024-01-01, 7d
    Run Garak Scans                :a2, after a1, 7d
    Reproduce 3 Handbook Attacks   :a3, after a2, 16d
    section Contribution
    First Bug Bounty Submission     :b1, after a3, 10d
    Contribute to OSS Tool          :b2, after b1, 10d
    Write Technical Blog Post       :b3, after b2, 10d
    section Career
    Build Custom Tool               :c1, after b3, 15d
    Apply to 5 Roles                :c2, after c1, 10d
    Interview Prep                  :c3, after c2, 5d

90-Day Action Plan Timeline

Days 1-30: Foundation Building

Week 1: Lab Setup

  • Set up isolated Ubuntu VM or Docker environment
  • Install Python 3.10+, transformers, torch
  • Clone and install Garak: pip install garak
  • Deploy local Llama-3-8B via Ollama
  • Test Burp Suite with LLM API proxy

Week 2: Automated Scanning

  • Run full Garak scan against local model
  • Analyze results, identify top 3 vulnerabilities
  • Manually verify each finding
  • Document findings in structured report

Week 3-4: Hands-On Attacks

  • Reproduce Prompt Injection (Chapter 14) against ChatGPT-like interface
  • Reproduce RAG Poisoning (Chapter 12) in local RAG setup
  • Reproduce Jailbreak (Chapter 16) using DAN or similar technique
  • Create GitHub repo documenting all 3 attacks with code

Deliverable: GitHub repository: my-ai-security-lab with working exploit code

Days 31-60: Real-World Application

Week 5-6: Bug Bounty

  • Read HackerOne/Bugcrowd policies for OpenAI, Google, Microsoft
  • Identify in-scope AI feature (ChatGPT plugins, Gemini Extensions, Copilot)
  • Test for prompt injection, data leakage, authentication bypass
  • Submit first report (even if it gets closed as duplicate - learning experience)

Week 7: Open Source Contribution

  • Pick tool: Garak, PyRIT, or NeMo-Guardrails
  • Find "good first issue" on GitHub
  • Submit PR (could be documentation, new probe, bug fix)
  • Engage with maintainers on review feedback

Week 8: Content Creation

  • Write 1,500+ word blog post on Medium/personal blog
  • Topic: "How I Found My First AI Vulnerability" or "Building a Custom LLM Fuzzer"
  • Include code, screenshots, diagrams
  • Share on Twitter, LinkedIn, relevant Discord servers

Deliverable: Published writeup + at least 1 bug bounty submission

Days 61-90: Career Launch

Week 9-10: Tool Development

  • Build custom tool addressing gap you've noticed:
    • Nuclei templates for AI endpoints
    • Automated jailbreak generator using GCG
    • Compliance mapper (findings → ISO 42001 controls)
    • RAG security scanner
  • Publish to GitHub with README, usage examples, tests
  • Add to portfolio as "Featured Project"

Week 11: Job Applications

  • Update LinkedIn: "AI Security Engineer" as target role
  • Update resume: Highlight handbook, projects, contributions
  • Apply to 5 companies:
    • 2 FAANG (Google, Microsoft, Meta AI safety teams)
    • 2 AI startups (OpenAI, Anthropic, Hugging Face, Cohere)
    • 1 security consultancy (Trail of Bits, NCC Group, Bishop Fox)

Week 12: Interview Preparation

  • Practice technical questions from Section 46.4
  • Prepare 3 "project deep dive" stories (tool you built, attack you found, defense you designed)
  • Do mock interviews with peers or use interviewing.io
  • Review handbook chapters on topics you're weakest in

Week 13: Networking

  • Attend 1 virtual conference or meetup (DEF CON Groups, AI Safety Summit online)
  • Connect with 10 AI security professionals on LinkedIn
  • Send informational interview requests to 3 practitioners
  • Join and actively participate in 1-2 relevant Discord/Slack communities

Deliverable: 5 job applications submitted + active interviews or next round scheduled

46.6 The Ethical Red Teamer's Oath

As you enter the field of AI security, consider adopting this professional code:

I pledge to:

Test Responsibly

  • Only test systems I am explicitly authorized to test
  • Respect the Rules of Engagement in every engagement
  • Stop immediately if I risk causing harm or accessing live customer data

Report Honorably

  • Disclose vulnerabilities responsibly through proper channels
  • Never weaponize findings or sell them to malicious actors
  • Give vendors reasonable time to patch before public disclosure

Protect Privacy

  • Treat all data encountered during testing as confidential
  • Never exfiltrate, store, or share PII beyond proof-of-concept
  • Securely delete evidence after the engagement concludes

Advance the Field

  • Share knowledge through responsible writeups, talks, and open source
  • Mentor newcomers and contribute to community resources
  • Advocate for security and safety in AI development

Stay Ethical

  • Decline work that serves surveillance, oppression, or harm
  • Question whether my actions serve the greater good
  • Remember that AI systems impact real people's lives

I understand that with the power to break AI systems comes the responsibility to make them safer.

Appendix A: Comprehensive Glossary

Term Definition
Adversarial Example An input specifically crafted to cause misclassification or unexpected model behavior.
Alignment The process of ensuring an AI system's goals match human values and intentions (via RLHF, Constitutional AI, etc.).
API Key Leakage Unintentional exposure of authentication credentials through model outputs or logs.
Attention Mechanism The core component of transformers that determines which tokens the model "pays attention to" when generating output.
Backdoor Attack Poisoning training data or model weights to create a hidden trigger that activates malicious behavior.
Chain of Thought (CoT) Prompting technique where the model shows its reasoning step-by-step, often improving accuracy but exposing logic.
Constitutional AI Anthropic's method of training models to follow ethical principles defined in a "constitution."
Context Window The maximum number of tokens (input + output) the model can process in a single interaction.
Data Poisoning Injecting malicious examples into training data to corrupt model behavior.
Embedding A dense vector representation of text/image/audio in high-dimensional space.
Few-Shot Learning Providing a few examples in the prompt to guide the model's behavior without fine-tuning.
Fine-Tuning Additional training on a specific dataset to specialize a pre-trained model.
Function Calling (Tool Use) The ability of an LLM to generate structured API calls to invoke external tools.
Gradient Descent The optimization algorithm used to train neural networks by minimizing loss.
Guardrails Input/output filters and safety mechanisms designed to prevent harmful model behavior.
Hallucination A confident but factually incorrect, fabricated, or nonsensical response.
Indirect Prompt Injection Injecting malicious instructions via external data sources (emails, web pages, documents) that the model reads.
Inference The process of using a trained model to generate predictions or outputs.
Jailbreak A prompt specifically designed to bypass safety training and content filters.
Logits Raw output scores from the model before applying softmax to convert to probabilities.
Mechanistic Interpretability Research field studying the internal representations and activations of neural networks to understand how they work.
Model Card Transparency document describing a model's intended use, training data, limitations, and evaluation results.
Model Extraction Stealing a model's functionality by querying it repeatedly and training a surrogate.
Model Inversion Recovering training data by analyzing model weights or outputs.
Perplexity A measure of how "surprised" a model is by a sequence; lower is better (more confident).
Prompt Injection Inserting malicious instructions that override or manipulate the System Prompt.
RAG (Retrieval-Augmented Generation) Enhancing a model's knowledge by retrieving relevant documents from an external database before generation.
RLHF (Reinforcement Learning from Human Feedback) Training technique where human raters score outputs to align model behavior with preferences.
Sponge Attack Input crafted to maximize computational cost, causing DoS through resource exhaustion.
System Prompt The initial hidden instructions given to the model by the developer (e.g., "You are a helpful assistant").
Temperature Sampling parameter controlling randomness. High (1.0+) = creative/unstable; Low (0.0-0.3) = deterministic /focused.
Token The basic unit of text the model processes (roughly 0.75 words in English).
Tokenization The process of converting text into tokens using algorithms like BPE (Byte-Pair Encoding).
Top-P (Nucleus Sampling) Sampling method that considers tokens with cumulative probability mass P.
Vector Database Database optimized for storing and searching high-dimensional embeddings (used in RAG).
Weight Poisoning Directly manipulating model parameters to introduce backdoors or degrade performance.
Zero-Shot Learning Asking the model to perform a task without providing any examples in the prompt.

Appendix B: The Red Teamer's Library (Essential Papers)

Foundational Research

  1. "Attention Is All You Need" (Vaswani et al., 2017)
    arXiv:1706.03762
    The transformer paper. Understand the architecture you're attacking.

  2. "Language Models are Few-Shot Learners" (GPT-3) (Brown et al., 2020)
    arXiv:2005.14165
    Demonstrated emergent capabilities and in-context learning.

  3. "BERT: Pre-training of Deep Bidirectional Transformers" (Devlin et al., 2018)
    arXiv:1810.04805
    Bidirectional understanding and the foundation for many NLP attacks.

  4. "Training language models to follow instructions with human feedback" (Ouyang et al., 2022)
    arXiv:2203.02155
    InstructGPT / ChatGPT - how RLHF creates alignment (and attack surface).

  5. "Constitutional AI: Harmlessness from AI Feedback" (Bai et al., 2022)
    arXiv:2212.08073
    Anthropic's alternative to RLHF and how constitutions can be attacked.

Attack Techniques

  1. "Universal and Transferable Adversarial Attacks on Aligned Language Models" (GCG) (Zou et al., 2023)
    arXiv:2307.15043
    The GCG algorithm - gradient-based jailbreak optimization.

  2. "Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection" (Greshake et al., 2023)
    arXiv:2302.12173
    The indirect injection paper that changed how we think about RAG security.

  3. "Extracting Training Data from Large Language Models" (Carlini et al., 2021)
    arXiv:2012.07805
    First demonstrated extraction of memorized training data from GPT-2.

  4. "Stealing Part of a Production Language Model" (Carlini et al., 2024)
    arXiv:2403.06634
    Model extraction from ChatGPT showing practical theft techniques.

  5. "Jailbroken: How Does LLM Safety Training Fail?" (Wei et al., 2023)
    arXiv:2307.02483
    Taxonomy of jailbreak techniques and why safety training fails.

  6. "Poisoning Web-Scale Training Datasets is Practical" (Carlini et al., 2023)
    arXiv:2302.10149
    Demonstrated how to poison LAION-400M at scale for $60.

  7. "Prompt Injection attack against LLM-integrated Applications" (Liu et al., 2023)
    arXiv:2306.05499
    Comprehensive study of injection vectors in production systems.

Defense and Robustness

  1. "Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned" (Ganguli et al., Anthropic, 2022)
    arXiv:2209.07858
    Methodology for responsible red teaming at scale.

  2. "Defending Against Backdoor Attacks in Natural Language Generation" (Yan et al., 2023)
    arXiv:2106.01022
    Detection and mitigation strategies for trojan attacks.

  3. "Certified Robustness to Adversarial Examples with Differential Privacy" (Lecuyer et al., 2019)
    arXiv:1802.03471
    Using DP for provable defenses.

  4. "Holistic Safety for Large Language Models" (Qu et al., 2024)
    arXiv:2311.11824
    Comprehensive framework for LLM safety evaluation.

  5. "NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications" (NVIDIA, 2023)
    Technical Report - GitHub
    Production defense architecture and implementation.

Advanced Topics

  1. "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" (Hubinger et al., 2024)
    arXiv:2401.05566
    Backdoors that survive RLHF - implications for alignment.

  2. "Tree of Attacks: Jailbreaking Black-Box LLMs Automatically" (Mehrotra et al., 2023)
    arXiv:2312.02119
    TAP algorithm - automated adversarial prompt generation.

  3. "The Adversarial Robustness of LLMs: A Survey" (Wang et al., 2024)
    arXiv:2407.12321
    Comprehensive survey of attack and defense landscape.

Compliance and Standards

  1. "Model Cards for Model Reporting" (Mitchell et al., 2019)
    arXiv:1810.03993
    Transparency framework - basis for compliance documentation.

  2. "Datasheets for Datasets" (Gebru et al., 2021)
    arXiv:1803.09010
    Data documentation for compliance and reproducibility.

  3. "NIST AI Risk Management Framework" (NIST, 2023)
    Technical Standard - nist.gov/itl/ai-risk-management-framework
    Official guidance for AI risk assessment.

Appendix C: Comprehensive Tool Repository

Vulnerability Scanners

Tool Install Use Case Quick Start
Garak pip install garak Automated LLM vulnerability scanning garak --model_name openai --model_type openai --probes all
PyRIT pip install pyrit Microsoft's Red Team automation framework python -m pyrit.orchestrator --target chatgpt
Inspect inspect-ai.org UK AI Safety Institute evaluation platform pip install inspect-ai; inspect eval
PromptFuzz GitHub:microsoft/promptfuzz Prompt fuzzing and testing Clone repo, follow README

Traffic Analysis and Proxies

Tool Install Use Case Quick Start
Burp Suite Download Intercept and modify LLM API calls Configure proxy, install Logger++ extension
mitmproxy pip install mitmproxy Programmable MITM for API traffic mitmproxy -s inject_script.py
Proxyman Download MacOS alternative to Burp Install, configure system proxy

Attack Frameworks

Tool Install Use Case Quick Start
TextAttack pip install textattack Adversarial text generation textattack attack --model bert-base-uncased --dataset ag_news
ART (IBM) pip install adversarial-robustness-toolbox Multi-framework adversarial attacks from art.attacks import FastGradientMethod
CleverHans pip install cleverhans Classic adversarial ML library from cleverhans.attacks import FastGradientMethod

Model Analysis and Interpretability

Tool Install Use Case Quick Start
TransformerLens pip install transformer-lens Mechanistic interpretability from transformer_lens import HookedTransformer
Ecco pip install ecco Explore what activates in transformers import ecco; lm = ecco.from_pretrained('gpt2')
BertViz pip install bertviz Visualize attention patterns from bertviz import head_view

Defense and Guardrails

Tool Install Use Case Quick Start
NeMo Guardrails pip install nemoguardrails NVIDIA's LLM control framework nemoguardrails --config config.yml
Guardrails AI pip install guardrails-ai Input/output validation from guardrails import Guard
Presidio pip install presidio-analyzer Microsoft PII detection and redaction from presidio_analyzer import AnalyzerEngine
LiteLLM pip install litellm Unified LLM API with filtering from litellm import completion

Specialized Reconnaissance

Tool Install Use Case Quick Start
Nuclei go install -v github.com/projectdiscovery/nuclei/v2/cmd/nuclei@latest Vulnerability scanner (AI templates) nuclei -t ai-templates/ -l targets.txt
httpx go install -v github.com/projectdiscovery/httpx/cmd/httpx@latest HTTP probing to find AI endpoints cat domains.txt | httpx -td
Subfinder go install -v github.com/projectdiscovery/subfinder/v2/cmd/subfinder@latest Subdomain enumeration subfinder -d target.com

Local Model Deployment (for testing)

Tool Install Use Case Quick Start
Ollama ollama.ai/download Run Llama/Mistral locally ollama run llama3
LM Studio lmstudio.ai GUI for local models Download, load model, start server
vLLM pip install vllm High-performance inference python -m vllm.entrypoints.api_server --model meta-llama/Llama-2-7b
Text Generation WebUI GitHub:oobabooga/text-generation-webui Gradio UI for testing Clone, install, run python server.py

Monitoring and Logging

Tool Install Use Case Quick Start
LangSmith Sign up at smith.langchain.com LangChain observability Set LANGCHAIN_TRACING_V2=true env var
Weights & Biases pip install wandb ML experiment tracking wandb login; wandb init
MLflow pip install mlflow Model lifecycle management mlflow ui

Appendix D: Quick Reference Cheat Sheet

OWASP Top 10 for LLM Applications

  1. LLM01: Prompt Injection - Manipulating LLM via crafted inputs
  2. LLM02: Insecure Output Handling - Insufficient output validation leading to XSS/injection
  3. LLM03: Training Data Poisoning - Compromising training data to create backdoors
  4. LLM04: Model Denial of Service - Resource exhaustion through expensive inputs
  5. LLM05: Supply Chain Vulnerabilities - Compromised models, datasets, or frameworks
  6. LLM06: Sensitive Information Disclosure - Leaking PII through model outputs
  7. LLM07: Insecure Plugin Design - Flaws in LLM extensions and tool integration
  8. LLM08: Excessive Agency - Over-privileged LLM actions without oversight
  9. LLM09: Overreliance - Trusting LLM outputs without verification
  10. LLM10: Model Theft - Unauthorized extraction or replication of models

Top 5 Attack Patterns (Critical)

  1. Indirect Prompt Injection via RAG

    • Poison documents in vector database
    • Wait for retrieval to inject malicious instructions
    • Model executes attacker's commands

  2. Function-Calling Privilege Escalation

    • Trick LLM into calling admin-only functions
    • Bypass intended access control logic
    • Achieve unauthorized actions

  3. Training Data Extraction

    • Craft prompts that trigger memorization
    • Extract PII, secrets, proprietary data
    • Verify with divergence metrics

  4. Multi-Turn Jailbreak

    • Build up context over multiple exchanges
    • Gradually erode safety alignment
    • Finally request harmful content

  5. Supply Chain Pickle RCE

    • Craft malicious PyTorch model file
    • Upload to model hub / send to victim
    • Arbitrary code execution on torch.load()

Defense-in-Depth Checklist

Input Layer

  • Unicode normalization (NFKC)
  • Invisible character stripping
  • Prompt injection detection classifier
  • Rate limiting by tokens (not requests)
  • API key rotation and scoping

Model Layer

  • Updated model with latest safety training
  • System prompt hardening (delimiters, explicit rules)
  • Temperature/top-p constraints for determinism
  • Function-calling ACLs (principle of least privilege)

Output Layer

  • PII redaction (Presidio or regex)
  • Harmful content filtering
  • Citation/grounding enforcement
  • Output length limits

Infrastructure

  • Model signing and provenance tracking
  • Dependency pinning (requirements.txt with hashes)
  • Secrets management (never in code)
  • Comprehensive logging (input/output/metadata)
  • Isolated execution environment for code generation

Monitoring

  • Real-time safety violation alerts
  • Token velocity dashboards
  • Anomaly detection on prompt patterns
  • Regular penetration testing (quarterly)
  • Bug bounty program

AI Security Incident Response

  1. Detect: Alert triggered (PII leak, jailbreak detected, unusual token volume)
  2. Contain: Circuit breaker activates, feature flagged off, isolate affected users
  3. Analyze: Pull logs, reproduce attack, assess scope of compromise
  4. Eradicate: Patch vulnerability, update guardrails, retrain if necessary
  5. Recover: Gradual rollout with enhanced monitoring, validate fix
  6. Learn: Post-mortem, update runbooks, add regression test

Appendix E: Chapter Cross-Reference Matrix

Attack/Topic Primary Chapter Related Chapters Tools Referenced
Prompt Injection 14 11, 18, 21 Garak, PyRIT, TextAttack
RAG Security 12 13, 18 ChromaDB, Pinecone, Weaviate
Jailbreaking 16 14, 19, 43 Gandalf, HackThePrompt
Model Theft 20 13, 25 ART, CleverHans
Data Poisoning 19 13, 26 TextAttack, ART
Plugin Exploitation 17 11, 23 Burp Suite, Nuclei
Model DoS 21 10, 41 Garak (sponge probes)
Multimodal Attacks 22 9, 44 Adversarial Robustness Toolbox
Social Engineering 23 2, 3 (Human tactics)
Adversarial ML 25 9, 18, 20 CleverHans, ART, TextAttack
Supply Chain 13, 26 7, 44 Picklescan, Sigstore
Federated Learning 27 25, 29 PySyft, Flower
Privacy Attacks 28 14, 15, 29 Membership Inference libraries
Model Inversion 29 20, 28 Research implementations
Backdoors 30 13, 19, 43 TrojanNN, BadNets replicas
OSINT Recon 31 6, 39 Shodan, Censys, theHarvester
Attack Automation 32 38, 43 Garak, PyRIT, Custom scripts
Defense Evasion 34 14, 41 Encoding tools, obfuscation
Reporting 36 2, 40 Markdown, LaTeX, CVSS calculators
Compliance 40 2, 5, 41 ISO 42001 templates, NIST RMF
Bug Bounties 39 14, 17, 36 Burp Suite, Nuclei, Garak
Building Programs 45 3, 38, 40 (Organizational tools)

Final Words

You've completed a journey through 46 chapters, over 1,000 pages, and the entire landscape of AI security. From the philosophical foundations of ethics and mindset to the cutting edge of algorithmic attacks and formal verification, you now possess a comprehensive understanding of how AI systems fail—and how to make them safer.

This is not the end. It's the beginning.

AI security is the most consequential field in cybersecurity today. The systems you test will make critical decisions: approving loans, diagnosing diseases, controlling autonomous vehicles, summarizing legal contracts, and moderating what billions of people see online. When these systems fail, the consequences are measured not just in dollars, but in lives, liberty, and the fabric of democratic discourse.

Your role as a Red Teamer is to find the failures before they matter.

The attacks you've learned—prompt injection, model extraction, RAG poisoning, jailbreaks—are not theoretical exercises. They're being deployed right now, in the wild, by adversaries ranging from curious hobbyists to nation-states. Your job is to think like them, but act in service of defense.

Remember:

  • Every system you harden makes the world slightly safer
  • Every vulnerability you report saves someone from exploitation
  • Every tool you build multiplies the effectiveness of defenders
  • Every person you mentor creates a network of protection

AI is not slowing down. GPT-5, Llama-4, Gemini-2, Claude-4—each more powerful, each with new capabilities, each with new attack surfaces. The red team never rests, because the models never stop evolving.

But now you're ready.

You know how transformers work at the token level. You can fingerprint AI backends, craft gradient-optimized jailbreaks, poison vector databases, and map findings to ISO 42001 controls. You can build tools, hunt bounties, and design compliance programs. You can communicate with executives, negotiate with clients, and mentor junior teammates.

The handbook may be complete, but your journey is just starting.

In 90 days, you could have deployed your first custom fuzzer. In 6 months, you could have submitted your first CVE. In a year, you could be leading a red team program at a major AI lab. In 5 years, you could be defining the standards that the next generation of practitioners follows.

AI is the API to human knowledge. Securing it is the prerequisite for a safe future.

Go forth. Build things. Break things responsibly. Find the bugs before they become disasters. Share what you learn. And remember that every test you run, every report you write, every system you secure contributes to a world where AI amplifies human potential instead of human vulnerability.

Welcome to the future. Let's make it secure.

— The AI LLM Red Team Handbook Team
Version 1.46.154 | Gold Master | January 2026

Contribute: github.com/Shiva108/ai-llm-red-team-handbook
License: CC BY-SA 4.0

Reference in New Issue View Git Blame Copy Permalink