ISC2 AI Security Pilot Exam · Free study guide
ISC2 AI Security Pilot Exam: A Free-First Study Guide
About the exam
The pilot exam for ISC2's new AI security certification runs from mid-December 2026 to January 24, 2027, at Pearson VUE test centers, and it's free for the up to 3,000 volunteers selected. It has 160 questions and a three-hour limit that includes breaks, and you can't skip a question and come back to it. Results arrive 8 to 12 weeks after the pilot closes, as a pass or fail with no domain scores. ISC2 hasn't announced the certification's name yet (ISC2 FAQ).
The outline has six domains. ISC2 hasn't published domain weights, so plan to study all six.
| Domain | What it covers |
|---|---|
| 1. AI Concepts and Security Fundamentals | ML and GenAI basics, limits like hallucination and drift, core AI threats and controls |
| 2. AI Governance, Ethics, Risk and Compliance | Frameworks and regulation, responsible AI, risk assessment, third-party risk, human oversight and audit |
| 3. AI Data Management, Privacy and Model Security | Data lineage, privacy techniques, training data integrity, model lifecycle, model theft and tampering |
| 4. Secure AI Architecture, Engineering and IAM | Secure design, infrastructure, MLOps pipelines, guardrails, human and non-human identities, agent permissions |
| 5. Secure AI Operations and Incident Response | AI telemetry, runtime detection, AI vulnerability management, containment, forensics, recovery |
| 6. AI Threats, Testing and Lifecycle Management | Threat modeling (STRIDE, PASTA, MAESTRO), AI supply chain, red teaming, reporting, mitigations |
Read the official outline itself before anything else; every section lists specific tasks, and those tasks are your real syllabus.
Where a CISSP helps, and where it doesn't
You don't need a CISSP for this exam. If you have one, it covers a real share of Domains 4 and 5 and parts of the others, but the AI-specific material in the right column is new for everyone. If you're studying for the IAPP AIGP, it overlaps heavily with Domain 2.
| Domain | Where a CISSP helps | New for everyone |
|---|---|---|
| 1. Fundamentals | Security principles, zero trust | ML and GenAI concepts, AI-specific attacks |
| 2. Governance and risk | Risk management, policy, third-party risk | EU AI Act, NIST AI RMF, responsible AI, human oversight of AI |
| 3. Data and model security | Data classification, encryption, key management | Poisoning, model theft, privacy-preserving ML, model lifecycle |
| 4. Architecture and IAM | Secure design, IAM, physical security | Agents, MCP and A2A, ML pipelines, non-human identities |
| 5. Operations and IR | Incident response, BC/DR, logging and monitoring | AI telemetry, containing a model or agent, model forensics |
| 6. Threats and testing | Threat modeling with STRIDE, secure SDLC, supply chain | MAESTRO, adversarial testing, AI red teaming, AI-BOMs |
The reference list matters more than it looks. ISC2's volunteer item writers check each question against reference sources, so the list shows where the answer key comes from. When a free framework and a reference book describe something differently, lean toward the book. ISC2 also says the list isn't all-inclusive, and other sources that cover the outline are fair game, which is what makes a free-first plan workable.
Study notes for every domain
I also wrote original study notes for each of the six domains, built from the ISC2 outline and the free frameworks below, so you can study without buying a single book. Each one follows the outline section by section, defines every term for someone starting from scratch, ends each section with what to expect on the exam, and closes with 12 ISC2-style practice questions and an answer key. Domains 2 through 6 also include diagrams.
| Domain | Study notes |
|---|---|
| 1. AI Concepts and Security Fundamentals | Domain 1 Study Notes |
| 2. AI Governance, Ethics, Risk and Compliance | Domain 2 Study Notes |
| 3. AI Data Management, Privacy and Model Security | Domain 3 Study Notes |
| 4. Secure AI Architecture, Engineering and IAM | Domain 4 Study Notes |
| 5. Secure AI Operations and Incident Response | Domain 5 Study Notes |
| 6. AI Threats, Testing and Lifecycle Management | Domain 6 Study Notes |
Use the notes as your main reading, the free core below for depth, and the book chapters if you have access.
Getting the books without a subscription
Most of the reference books are on O'Reilly's learning platform, and your public library card is the cheapest way in. Many US and Canadian library systems subscribe to O'Reilly for Public Libraries, including San Francisco, San Jose, Charlotte Mecklenburg and Knox County. Search your library's website for "O'Reilly" and sign in with your card number. The library version doesn't include every feature of a paid plan, but the books are the same.
If your library doesn't carry it, here are other routes:
- ACM membership. ACM Professional Members get O'Reilly through the ACM Skills Bundle add-on, which is an extra cost on top of membership.
- Your employer or school. Many companies and universities already pay for O'Reilly. University libraries also carry titles through ProQuest Ebook Central, and some of these books show up on OverDrive, which you can borrow through Libby.
- Interlibrary loan. Two references aren't on O'Reilly: Combating Cyberattacks Targeting the AI Ecosystem (Sood) and Security Policies and Implementation Issues (Johnson and Easttom). Ask your library to request them. The second one covers general security policy, so skip it if you already know that ground, whether from a CISSP or from experience.
You don't need all 23 books. Each domain section below names the free sources first, then the few chapters worth reading. The core list is eight books: Wilson, Sotiropoulos, Santos and Radanliev, Malik, Huang and Dawson, Sayles, Wendt, Ping, and Sood.
The free core
If you only have time for one free source, make it the OWASP AI Exchange. It runs to more than 300 pages of threats and controls across the whole AI lifecycle, it's released under CC0, and it fed directly into ISO/IEC 27090 and the EU AI Act's security standard. Everything below fills in specific domains.
| Resource | What it gives you | Domains |
|---|---|---|
| OWASP AI Exchange | Threats and controls for the full AI lifecycle, development through runtime | All |
| OWASP Top 10 for LLM Applications 2026 | The ten biggest LLM app risks, now ranked with real incident data, mapped to MITRE ATLAS | 1, 6 |
| OWASP Top 10 for Agentic Applications 2026 | Agent risks like tool misuse, goal hijacking and identity abuse | 4, 6 |
| OWASP Agentic AI Threats and Mitigations | A reference architecture for agents and the threat model behind the list above | 4, 6 |
| MITRE ATLAS | Adversary tactics and techniques against AI systems, built like ATT&CK | 1, 5, 6 |
| NIST AI 100-2 E2025 | NIST's taxonomy of attacks: evasion, poisoning, privacy attacks, misuse, plus mitigations | 1, 3, 6 |
| NIST AI RMF and the GenAI Profile (AI 600-1) | Govern, Map, Measure, Manage, and the GenAI-specific risks | 2 |
| EU AI Act | Risk tiers, prohibited uses, high-risk obligations | 2 |
| AI Data Security (CISA, NSA, FBI) | Data supply chain risk, poisoned data and data drift, with ten practice areas | 3 |
| Guidelines for Secure AI System Development (NCSC, CISA) | Secure design, development, deployment and operation | 4, 5 |
| CSA MAESTRO | A seven-layer threat model for agentic AI, compared against STRIDE and PASTA | 6 |
| CSA Agentic AI Red Teaming Guide | Test categories, steps and example prompts for red teaming agents | 6 |
| Google Machine Learning Crash Course | ML fundamentals for anyone without a data science background | 1 |
| ISC2 Code of Ethics | The canons; ISC2 exams always test them | All |
The outline names MITRE ATLAS, the OWASP lists, the NIST AI RMF, the EU AI Act, STRIDE, PASTA and MAESTRO directly, so expect questions that assume you know their vocabulary.
Domain 1: AI Concepts and Security Fundamentals
This is the vocabulary domain, and every other domain assumes it. If you don't come from data science, start here.
Know these cold
- ML types (supervised, unsupervised, reinforcement), deep learning, and the CNN, RNN and transformer architectures, plus what GPUs and TPUs do.
- GenAI building blocks: LLMs, multimodal models, prompt engineering, RAG, fine-tuning with LoRA and PEFT, embeddings, tokens and context windows, vector databases, agents and agent memory.
- Limits: hallucination, non-determinism, model and data drift, model collapse, evaluation metrics, and when to keep a human in the loop.
- Data basics: provenance, labeling and quality, plus tokenization, masking, k-anonymity and synthetic data.
- The core threats by name: prompt injection, jailbreaking, evasion, exfiltration, model inversion and data poisoning, along with zero trust applied to AI.
- Guardrails: input and output validation, toxicity filtering, model verification and isolating AI environments.
Free sources: Google ML Crash Course for the fundamentals, then the OWASP LLM Top 10 2026, the taxonomy in NIST AI 100-2 and the MITRE ATLAS matrix.
Book chapters: Ping, Ch 1, 3, 12–13 and 15–16; Santos and Radanliev, Ch 2–5; Wilson, Ch 2, 4–7 and 12; Malik, Huang and Dawson, Ch 1–2; Sotiropoulos, Ch 3.
Study notes: Domain 1 Study Notes, with 12 practice questions.
Domain 2: AI Governance, Ethics, Risk and Compliance
If you hold or are studying for the IAPP AIGP, you have a head start here. The difference is that this exam frames governance from a security leader's seat.
Know these cold
- Frameworks and regulation: the EU AI Act, NIST AI RMF, ISO/IEC 42001 and GDPR; how AI governance differs from IT governance; AI policies, use-case inventories, and roles like Chief AI Officer and AI ethics officer.
- Responsible AI principles (fairness, reliability, safety, privacy, inclusiveness, transparency), bias mitigation across protected groups, and explainability.
- Risk: AI impact assessments that weigh blast radius and autonomous actions, risk tiers, risk treatments, and what triggers a reassessment, such as a model update or drift.
- Privacy and IP: sensitive data types, copyright and fair use of training data, retention and deletion, and consent.
- Third parties: vendor due diligence, contract terms like security SLAs and incident notification, SBOMs and AI-BOMs, and shared responsibility.
- Oversight: human-in-the-loop versus human-on-the-loop, override authority, shadow AI detection, labeling AI-generated content, audits for bias and drift, and role-based training for AI operators.
Free sources: the NIST AI RMF and its GenAI Profile, the EU AI Act, the governance sections of the OWASP AI Exchange and the ISC2 Code of Ethics.
Book chapters: Sayles, Ch 4, 6–8, 10–12, 17–19 and 21; Wendt, Ch 3, 5 and 6; Santos and Radanliev, Ch 7–8; Crossley's chapter on third-party suppliers.
Study notes: Domain 2 Study Notes, with diagrams and 12 practice questions.
Domain 3: AI Data Management, Privacy and Model Security
For most security people this is the weakest domain, because it's about the data and the model themselves, not the infrastructure around them. Give it extra time.
Know these cold
- Data governance for AI: owner, steward and controller roles, residency and sovereignty, classification and tagging, lineage, chain of custody, and real versus synthetic data.
- Privacy-preserving techniques: differential privacy, anonymization, tokenization, noise injection and federated learning, and what a right-to-erasure request means for a model that's already trained.
- Attacks on training data: poisoning, backdoors, label tampering and malicious datasets, plus dataset versioning and the risks of training on synthetic output.
- Protection: DSPM, DLP and semantic filtering of prompts and outputs, encryption at rest, in transit and in use, key management, and crypto erase.
- Model lifecycle: ownership, risk classification, approval workflows, model cards, approved registries, rollback, and drift and bias monitoring.
- Model confidentiality, integrity and availability: protecting weights, preventing extraction, watermarking and fingerprinting, signing and hashing, rate limits and quotas.
- Supply chain: third-party and open datasets, provenance checks, and the risks of foundation and pretrained models.
Free sources: the joint AI Data Security guidance from CISA, NSA and FBI, the poisoning and privacy attack sections of NIST AI 100-2, and the development-time threats in the OWASP AI Exchange.
Book chapters: Sotiropoulos, Ch 4–6, 8, 10 and 15; Sayles, Ch 4, 19 and 21; Santos and Radanliev, Ch 7; Wilson, Ch 5 and 8; and Nelson's Is Building Secure ML Possible?, a 24-page report you can finish in an hour.
Study notes: Domain 3 Study Notes, with diagrams and 12 practice questions.
Domain 4: Secure AI Architecture, Engineering and IAM
If you've worked in security architecture or IAM, much of this domain will feel familiar, and a CISSP covers a good part of it. The new material for everyone is agents and ML pipelines.
Know these cold
- Secure design for AI: zero trust, defense in depth and secure defaults; threat modeling RAG, vector databases and inference endpoints; build versus buy.
- Integrations: securing MCP servers, agent-to-agent (A2A) protocols, plugins and APIs.
- Infrastructure: hardening cloud, on-prem and hybrid AI; sandboxing training and inference; compute throttling; isolated accelerator capacity; physical and environmental controls.
- Pipelines (MLOps and LLMOps): secure CI/CD and containers, ingestion integrity, lineage for fine-tuning, validation gates, and data unlearning.
- Guardrails: prompt sanitization, content filtering, context boundary checks, and guardrails at the orchestration layer for autonomous agents.
- Identity: human versus non-human identities (agents, workloads, tools), machine-to-machine auth, API tokens and secrets management.
- Authorization: least privilege and separation of duties across agents, RBAC, ABAC, PBAC and just-in-time access, execution boundaries, and the risk of excessive agency and over-delegation.
- Lifecycle: inventories of human and non-human identities, provisioning and deprovisioning, access reviews, token accounting and session timeouts.
Free sources: the NCSC and CISA guidelines for secure design and deployment, the OWASP Top 10 for Agentic Applications, and OWASP Agentic AI Threats and Mitigations. The OWASP GenAI Security Project also publishes a practical guide to secure MCP server development.
Book chapters: Malik, Huang and Dawson, Ch 4, 8 and 10–12, plus the appendix on agentic risks; Wilson, Ch 3, 4, 7, 8 and 11; Ping, Ch 7; Sotiropoulos, Ch 18; Santos and Radanliev, Ch 6.
Study notes: Domain 4 Study Notes, with diagrams and 12 practice questions.
Domain 5: Secure AI Operations and Incident Response
Start from the incident response lifecycle you already know, then learn what changes when the asset is a model or an agent.
Know these cold
- Readiness: model owner and data owner roles, release criteria, AI-specific playbooks and tabletops for poisoning and tampering, and configuration drift.
- Telemetry: logging prompts, completions, embeddings, tool calls and API traffic; correlating them in a SIEM or SOAR; keeping logs tamper-evident; masking sensitive data in logs.
- Detection: misuse across inputs, retrieval chains and outputs; over-privileged agents; exposed secrets; odd tool execution; UEBA and threat hunting.
- Vulnerability management: registers of model vulnerabilities, prioritizing by context and consequence, patching pipelines and dependencies, and compensating controls for legacy models.
- Response: an AI incident taxonomy (prompt injection, jailbreak, poisoning, model theft), escalation across SecOps, MLOps, data science and legal, and containment by rollback, isolation, rate limits, pipeline suspension or terminating an agent. Preserve weights, configs and dataset snapshots as evidence.
- Forensics and recovery: reconstructing the kill chain, root cause analysis, eradication by cleaning data and retraining, regression testing before reintroduction, clean checkpoints, vector store recovery, kill switches, BC/DR with AI dependency mapping, and decommissioning with machine unlearning.
Free sources: the operations section of the NCSC and CISA guidelines, the case studies in MITRE ATLAS, and the runtime sections of the OWASP AI Exchange.
Book chapters: Malik, Huang and Dawson, Ch 12; Wendt, Ch 8–9; Santos and Radanliev, Ch 6; Sood, Ch 3, 4 and 6; Wilson, Ch 11; Crossley for vulnerability management.
Study notes: Domain 5 Study Notes, with diagrams and 12 practice questions.
Domain 6: AI Threats, Testing and Lifecycle Management
This is the attacker's side of the exam: how to model threats against AI, test for them, and turn findings into fixes. Study it early, because the defensive domains make more sense once you know what they defend against.
Know these cold
- Threat modeling: STRIDE, PASTA and MAESTRO applied to AI pipelines and agents, human and non-human threat actors, insiders, and compute and token exhaustion attacks.
- Attack classes: evasion, poisoning, extraction, inversion and backdoors, plus threats specific to GenAI, RAG, agents and multimodal models.
- Secure development lifecycle: environment separation, versioning models, datasets and retrieval sources, canary deployments and rollback.
- Supply chain and procurement: vendor certifications and assessment reports, licensing and data ownership, hashing and signatures for provenance, and AI-BOMs.
- Testing: scoping an AI security assessment, adversarial testing of GenAI, RAG and agents, red team methods, and AI-assisted testing tools.
- Findings: triage by severity, exploitability, reproducibility and business impact; telling a real vulnerability from a model limitation like hallucination; mapping to MITRE ATLAS and the OWASP lists; reporting to technical and executive audiences.
- Mitigations: adversarial training, provenance tracking, data validation, filtering, sandboxing, rate limits and watermarking, then regression testing and monitoring residual risk.
Free sources: CSA MAESTRO, the CSA Agentic AI Red Teaming Guide, MITRE ATLAS, the OWASP LLM Top 10 2026 and NIST AI 100-2.
Book chapters: Sotiropoulos, Ch 3, 7, 14 and 17–19; Sood, Ch 5; Santos and Radanliev, Ch 5–6; Wilson, Ch 2, 8, 9 and 11; Malik, Huang and Dawson, Ch 7 and 9.
Study notes: Domain 6 Study Notes, with diagrams and 12 practice questions.
A 9-week plan
At four to six hours a week, nine weeks is enough. Count back nine weeks from your exam date to find your Week 1; with a January 24 deadline, that means starting by late November. The plan puts the attacker's view (Domain 6) early and saves governance (Domain 2) for late, since many candidates already know it.
| Week | Focus | Start with |
|---|---|---|
| 1 | Domain 1: ML and GenAI fundamentals | Google ML Crash Course; Ping, Ch 1 and 3 |
| 2 | Domain 1: threats and controls, plus ethics | OWASP LLM Top 10 2026; ISC2 Code of Ethics; Wilson, Ch 2 and 4 |
| 3 | Domain 6: threat modeling and testing | MAESTRO, MITRE ATLAS, NIST AI 100-2; Sotiropoulos, Ch 3 and 7 |
| 4 | Domain 3: data and model security | CISA AI Data Security; Sotiropoulos, Ch 4–6 and 8 |
| 5 | Domain 4: architecture, pipelines and agent identity | NCSC and CISA guidelines; OWASP Agentic Top 10; Malik, Huang and Dawson |
| 6 | Buffer week: catch up and review | Flashcards on your weakest sections |
| 7 | Domain 5: operations and incident response | MITRE ATLAS case studies; Malik, Huang and Dawson, Ch 12; Wilson, Ch 11 |
| 8 | Domain 2: governance, risk and compliance | NIST AI RMF, EU AI Act; Sayles; Wendt |
| 9 | Practice questions, re-rate every section, sit the exam | Santos and Radanliev's end-of-chapter exercises |
Week 6 is a buffer for the holidays or a week that slips. If you have fewer than nine weeks, drop it and move your weakest domain to the front. Before you start, rate yourself on every outline section as strong, shaky or new, and spend the extra time where the rating says "new."
Practice and exam day
There are no official practice tests yet, so you'll have to build your own practice.
- Start with the 72 questions in these notes. Each domain page ends with 12 ISC2-style questions and an answer key.
- Use the end-of-chapter exercises. Every chapter of Beyond the Algorithm (Santos and Radanliev) ends with "Test Your Skills" questions, which are the closest thing to a practice bank right now.
- Write questions from the outline. Turn each task statement into a scenario question with BEST, FIRST or MOST wording, then answer from a risk-and-governance seat; ISC2 exams reward the manager weighing risk over the technician fixing one system. An AI assistant can draft these for you if you give it the outline.
- Practice without going back. The exam is linear, so you can't skip a question and return to it later. With 160 questions in three hours, you have a little over a minute each, including breaks.
- Book early. Pearson slots between mid-December and January 24 overlap the holidays, and those fill fast.
- Respect the NDA. You'll agree to keep exam content confidential, so share how you studied, never what you saw. This guide uses only ISC2's public outline, its reference list, and free public frameworks.
- Expect a wait. Results come 8 to 12 weeks after the pilot window closes, as a pass or fail. The pilot is partly a test of the exam itself, so don't read too much into how any one question felt.
Good luck. If this helped, pass it along to someone else who's sitting the pilot.