Further Reading: AI and Machine Learning in Security
Curated, annotated resources to deepen this chapter. Each entry notes which learning path it serves most (🛡️ SOC, 🏗️ Engineer, 📋 GRC, 📜 Cert) and its citation tier. Start with the suggested order; you do not need to read everything before Chapter 35.
Suggested order
- Read the OWASP Top 10 for LLM Applications — short, concrete, and immediately useful before you let any LLM near your data or tools.
- Skim MITRE ATLAS to see adversarial-ML attacks (poisoning, evasion) organized the way ATT&CK organizes intrusions.
- Read the NIST AI Risk Management Framework overview for the governance lens your GRC team will need.
- For the statistics, revisit any solid treatment of anomaly detection and the base-rate fallacy — the two ideas that make or break a deployment.
Standards & primary documents (Tier 1)
- OWASP, Top 10 for Large Language Model (LLM) Applications (owasp.org). 🛡️🏗️📋 The community reference for LLM risks — prompt injection, insecure output handling, training-data poisoning, excessive agency, sensitive-information disclosure. Read this before deploying any LLM assistant in the SOC; it directly informs §34.5.
- MITRE, ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) (atlas.mitre.org). 🛡️🏗️ The ML-specific cousin of ATT&CK: a public knowledge base of real-world attacks on ML systems (poisoning, evasion, model theft) with case studies. The map for §34.4's threat model.
- NIST, AI Risk Management Framework (AI RMF 1.0) (NIST AI 100-1, 2023). 📋🏗️ The authoritative governance framework for trustworthy AI — Govern, Map, Measure, Manage. The board-level lens for AI in security, including the security-of-AI and AI-for-security distinction.
- NIST, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2). 🏗️ A rigorous taxonomy of evasion, poisoning, and privacy attacks and their defenses; the precise-vocabulary companion to §34.4.
- NIST SP 800-94, Guide to Intrusion Detection and Prevention Systems (IDPS). 🛡️ Frames signature- based versus anomaly-based detection at the standard level — the foundation this chapter generalizes to learned models.
- MITRE ATT&CK (attack.mitre.org). 🛡️ Revisit from Chapter 22: the behavioral techniques your anomaly detector and UEBA use cases should map to, so "unusual" connects to "which adversary technique."
Books & long-form (Tier 1 / Tier 2)
- Chio, C., & Freeman, D., Machine Learning and Security (O'Reilly). 🏗️🛡️ The practitioner's book on exactly this chapter's subject — anomaly detection, classifier evaluation, adversarial ML, and the operational realities (including the base-rate problem) of security ML. The single best deeper read.
- Sommer, R., & Paxson, V., "Outside the Closed World: On Using Machine Learning for Network Intrusion Detection" (IEEE S&P, 2010). 🛡️🏗️ The classic, still-relevant paper on why anomaly-based intrusion detection is so much harder than it looks — high cost of errors, the semantic gap, evaluation difficulties. Read it as the antidote to vendor hype. (Tier 2: a specific paper; confirm the venue if citing formally.)
- Goodfellow, I., Bengio, Y., & Courville, A., Deep Learning (MIT Press). 🏗️ Background on the models themselves for engineers who want to understand what is under the hood; dip into the relevant chapters rather than reading cover to cover.
- Anderson, R., Security Engineering (3rd ed.). 🏗️📋 The chapters on detection, fraud, and reasoning-about-attackers frame why a control's operational economics (false positives, who acts on alerts) matter more than its theoretical power — the §34.3 lesson in a broader setting.
Free online & talks (Tier 1 / Tier 2)
- CISA / NCSC joint guidance on AI and on Secure AI System Development. 📋🏗️ Government guidance on building and deploying AI securely, including secure design, data provenance, and the supply chain of models — the policy framing for §34.4's defenses. (Tier 2: title/issuer pattern; verify current document before citing.)
- Reporting on real deepfake-fraud incidents (2023–2024). 📋🛡️ Well-sourced accounts of finance employees tricked by deepfaked voices/video calls into authorizing multi-million-dollar transfers — the real pattern behind Case Study 2. (Tier 2: read reputable, well-sourced coverage; specifics vary by retelling.)
- The base-rate fallacy, any rigorous explainer. 📜🛡️ Understanding why a rare-event detector produces mostly false positives is the most important quantitative idea in this chapter; internalize it from any solid probability source. (Tier 2.)
Tools to explore (in your own authorized lab only)
- A hand-built z-score / UEBA notebook. 🛡️🏗️ Extend this chapter's
mlsec.py: collect a per-entity baseline from your own lab logs, compute z-scores, then add a second feature and fuse them — feel the jump from a single-feature score to a UEBA-style risk score (and the explainability cost). - An adversarial-examples toolkit (defensive study). 🏗️ Libraries exist to generate evasion examples against your own models so you can test robustness. Use them only on models you own or are authorized to test (§34.4 ethics note).
- A local/sandboxed LLM with deliberate prompt-injection test cases. 🛡️🏗️ Build a toy "summarize this email" assistant and try to hijack it with hidden instructions in the email body — the fastest way to internalize why least privilege and human-in-the-loop are non-negotiable.
⚖️ Authorization & Ethics reminder: Several resources describe how to attack ML systems (evasion, poisoning) and LLMs (prompt injection). Study them to defend; apply them only to systems you own or are explicitly authorized to test (Chapter 39). Crafting adversarial inputs against a third party's deployed model can cause real harm and is very likely illegal.