The short version
This is a review sheet, not a full course. It lists the facts and the term pairs that candidates most often mix up. It does not cover everything on the exam, and CompTIA’s current objectives are the authority. Exam facts were checked on CompTIA’s SecAI+ page on October 22, 2026.
Exam facts at a glance
| Item | Fact |
|---|---|
| Exam | CY0-001 (V1), CompTIA SecAI+ |
| Questions | Maximum of 60, multiple-choice and performance-based |
| Time | 60 minutes |
| Passing score | 600 on a scale of 100 to 900 |
| Languages | English and Japanese |
| Domains | 17% basics, 40% securing AI, 24% AI-assisted security, 19% governance |
| Renewal | 15 CEUs, three-year cycle |
Domain 1: Basic AI concepts (17%)
Key ideas
- Generative AI creates new content. Machine learning learns patterns from data. Deep learning uses many-layered neural networks.
- Large language models are broad and heavy. Small language models are narrower and cheaper to run and to protect.
- Supervised learning uses labeled data. Unsupervised learning finds structure without labels. Reinforcement learning learns from rewards.
- Fine-tuning changes the model with more training. Prompt engineering changes only the input.
- Retrieval-augmented generation (RAG) feeds retrieved documents to a model at question time, usually through embeddings and a vector store.
Term pairs
- Zero-shot has no examples in the prompt. One-shot has one. Multi-shot has several.
- Data lineage tracks where data traveled. Provenance records where it came from.
- Human-in-the-loop means a person is part of the process. Human oversight means a person supervises. Human validation means a person checks an output.
Decision rule. If the scenario says the model cannot be retrained, think prompts, retrieval, or guardrails, not fine-tuning.
Domain 2: Securing AI systems (40%)
Key ideas
- Threat-modeling resources include the OWASP lists for LLMs and machine learning, and MITRE ATLAS.
- Guardrails limit what a model will do. Prompt templates structure input. A prompt firewall filters traffic at a gateway.
- Rate limits, token limits, and quotas bound volume and cost.
- Protect data in transit, at rest, and in use.
Attack and control pairs
| If the scenario describes… | Think… |
|---|---|
| Crafted input that overrides instructions | Prompt injection: input validation, guardrails, least privilege for the model |
| Bad records added to training data | Data poisoning: provenance checks, data validation |
| A tampered model file | Model poisoning or supply chain: integrity checks, trusted sources |
| Recovering training data from outputs | Model inversion: limit outputs, privacy techniques |
| Testing whether a record was in training data | Membership inference: privacy techniques, access limits |
| Sensitive data in prompts or logs | Redaction, masking, log sanitization |
Confusable terms
- Masking hides values but keeps format. Redaction removes them. Anonymization removes the link to a person.
- Rate limit bounds volume. A guardrail shapes content. They solve different problems.
- Monitoring watches live behavior. Auditing reviews records afterward.
Decision rule. Start with the layer where the problem occurs: data, model, access, or gateway. Then pick the control that sits at that layer.
Domain 3: AI-assisted security (24%)
Key ideas
- Defenders use AI for anomaly detection, summarization, vulnerability analysis, fraud detection, and automating repetitive work.
- Attackers use it for deepfakes, impersonation, reconnaissance, social engineering, and obfuscation.
- Pipelines can add code scanning, software composition analysis, regression testing, and model testing.
Term pairs
- Anomaly detection flags departures from normal. Pattern recognition matches known shapes.
- Misinformation is false, not always deliberate. Disinformation is deliberately false.
- A low-code tool needs little scripting. A no-code tool needs none.
Decision rule. A new or unknown behavior points to anomaly detection. A known signature points to matching.
Domain 4: AI governance, risk, and compliance (19%)
Key ideas
- Governance covers policies, roles, and structures such as a central AI team.
- Responsible-AI principles include fairness, transparency, explainability, accountability, privacy, and inclusiveness.
- Named frameworks include the EU AI Act and the NIST AI Risk Management Framework.
- Shadow AI is unapproved AI use inside an organization.
Term pairs
- Transparency says AI is in use. Explainability gives the reason for one result.
- Policy states the rule. Procedure lists the steps.
- Bias is skewed outcomes. Data leakage is sensitive data escaping.
Decision rule. If the scenario is about a legal or regional requirement, look for the framework or data-sovereignty answer before a technical control.
Common traps
- Picking a real control that sits at the wrong layer.
- Missing the one constraint sentence, such as “data may not leave the region.”
- Skipping the words “choose two” in a multi-select item.
- Spending too long on one performance-based item. Flag it and return.
- Reading “best” as “most complete” when the scenario asks for “first.”
Last-day checklist
- Re-read your own list of attacks and controls once, without notes.
- Redo only the questions you got wrong in the past week.
- Check your booking, ID, and start time. The exam day guide has the routine.
- Remember: no phones, no AI tools, and notes only on materials provided.
- Stop studying early and sleep.
More guides for CY0-001
- CY0-001 Study Plan: 4 and 8 Weeks
- CY0-001 Sample Questions With Answers
- Is CY0-001 Hard? SecAI+ Difficulty
- CY0-001 Prerequisites and Skills Check
- CY0-001 Renewal and Retakes Explained
- Is CY0-001 Worth It? An Honest Look
- SY0-801 vs CY0-001: Which Exam First?
Next step
Use the domains explained page for depth and the exam format page for rules. From the CY0-001 overview you can reach every guide. To test the terms above under a clock, try our CY0-001 practice tests.