Knowledge is Power
BARE METAL CYBERKnowledge is Power

Sitewide Search

Search Bare Metal Cyber

Search exact control and technique identifiers, Cyber Wiki articles, framework records, playbooks, books, podcasts, Academy courses, individual lessons, and BMC Cyber Challenge activities.
Adversarial Machine Learning cover

Ai Security

Adversarial Machine Learning

Mechanisms, Vulnerabilities, and Strategies for Trustworthy AI

A professional reference on attacks against machine-learning systems and the practices used to evaluate and defend them.

Wiley2026
View on Amazon

Reading and application plan

Turn the book into a repeatable learning cycle.

Read with a question in mind, use the Cyber Wiki to expand unfamiliar concepts, and convert each major idea into an action, decision, or discussion point.

Five-step cycle

  1. Preview the section

    Scan the chapter or major subject area and write down the question you want the reading to answer.

  2. Read and annotate

    Capture the central claim, important terms, examples, and any decision you would make differently.

  3. Expand the concept

    Open two or three related Cyber Wiki articles to compare terminology, controls, threats, or implementation context.

  4. Apply the lesson

    Use the material in a real scenario, policy discussion, assessment, design review, class activity, or leadership decision.

  5. Review and retain

    Summarize the lesson in your own words and revisit the related concepts that remain uncertain.

This reading plan is Bare Metal Cyber editorial guidance and is separate from the publisher’s official book content.

Inside the book

Major subjects and chapter areas

The Age of Intelligent ThreatsAnatomy of AI Systems and Their Attack SurfacesThe Adversary’s PlaybookEvasion Attacks — Tricking AI Models at InferencePoisoning Attacks — Compromising AI Systems During TrainingPrivacy Attacks — Extracting Secrets from AI ModelsBackdoor and Trojan Attacks — Embedding Hidden Behaviors in AI ModelsThe Generative AI Attack SurfacePrompt Injection and Jailbreak TechniquesData Leakage and Model HallucinationAdversarial Fine-Tuning and Model ReprogrammingAgentic AI and Autonomous Threat Loops

Continue learning

Explore the related Cyber Wiki.

Direct and course-derived connections come from the existing Bare Metal Cyber catalog. Broader topic links are editorial recommendations, not publisher claims or official framework mappings.

Direct book connection

Security Assumptions

Security assumptions are conditions believed to be true about users, systems, threats, dependencies, and controls; they must be stated and tested because designs fail when hidden assumptions are wrong.

Open article →Related topic

Adversarial Machine Learning

Adversarial machine learning studies attacks and mitigations involving model training, inference, data, outputs, and operational dependencies.

Open article →Related topic

AI Feedback Loops

AI feedback loops occur when a system's outputs influence the future data used to evaluate or retrain it.

Open article →Related topic

Privacy-Preserving Machine Learning

Privacy-preserving machine learning applies technical and governance measures to limit exposure of personal or sensitive information during training, evaluation, inference, and sharing.

Open article →Related topic

Secure Fine-Tuning

Secure fine-tuning adapts a pretrained model while controlling training data, code, adapters, permissions, evaluation, and release lineage.

Open article →Related topic

AI System Incident Response

AI-system incident response adapts preparation, detection, analysis, containment, recovery, and learning to failures and attacks involving models, data, prompts, outputs, pipelines, and automated decisions.

Open article →Related topic

Inference API Security

Inference API security protects model-serving interfaces against unauthorized use, abuse, extraction, privacy attacks, resource exhaustion, and unsafe input or output handling.

Open article →Related topic

Prompt Injection and Agentic AI Risk

Prompt injection attempts to influence an AI system through untrusted instructions, while agentic risk grows when a model can use tools, retrieve data, or take actions beyond producing text.

Open article →Related topic

Machine-Learning Observability

Machine-learning observability combines system, data, model, and business evidence so teams can understand whether an AI service is healthy and trustworthy in operation.

Open article →Related topic

LLM Jailbreak Resistance

Jailbreak resistance is the ability of an AI application to maintain enforced restrictions despite persuasive, encoded, multi-turn, or role-play attempts to bypass them.

Open article →Related topic

Direct Prompt Injection

Direct prompt injection is an attempt by a user to override or manipulate an AI application's intended instructions through supplied input.

Open article →Related topic

Shadow AI Discovery and Response

Shadow AI discovery and response identifies unauthorized AI tools, models, agents, extensions, data flows, and automated uses, then brings them under governance or removes them.

Open article →