Red Teaming AI. A Field Manual for Attacking Intelligent Systems, Філіп А. Дерсі

Опис

AI is no longer a futuristic concept—it’s embedded in critical systems shaping finance, healthcare, infrastructure, and national security. But with this power comes unprecedented risk. Red Teaming AI arms you with the mindset, methodology, and tools to proactively test and secure intelligent systems before real adversaries exploit them.

Written for security professionals, researchers, and AI practitioners, this field manual goes beyond theory. You’ll learn how to map the new AI attack surface, anticipate adversarial moves, and simulate real-world threats to uncover hidden vulnerabilities.

You’ll Learn How To:

Think in graphs, not checklists: trace attack paths through interconnected AI components, data pipelines, and human interactions
Poison the well: explore how adversaries corrupt training data to implant backdoors and erode model integrity
Fool the oracle: craft evasion attacks that manipulate AI perception at decision time
Hijack conversations: execute prompt injection attacks that turn Large Language Models into insider threats
Steal the brain: probe for model extraction and privacy attacks that compromise valuable IP
Conduct full-spectrum campaigns: use the STRATEGEMS framework and the AI Kill Graph to plan, execute, and report professional-grade red team engagements

Traditional security methods can’t keep up with adversarial AI. From manipulated financial agents to compromised autonomous vehicles, real-world failures have already caused billions in losses and threatened lives. Red Teaming AI equips you to meet this challenge with practical techniques grounded in real attack scenarios and cutting-edge research.

Introduction

Part I: The Adversarial Playbook: Mindset & Methodology