X1. AI security

Outcome: you can threat-model, protect and monitor a small AI application, use AI in security work without trusting it blindly, and explain AI risk and compliance in business terms.

Before you start: S2, S6, S7 and S11, and Security+ or equivalent knowledge. A little Python helps in lab 1; an AI assistant can draft the script, and you read every line before you run it.

Skills: AI types and how models are trained and tuned (machine learning, deep learning, generative AI, large and small language models); prompt engineering (system and user prompts, zero-, one- and multi-shot); data security for AI (cleansing, lineage, provenance, vector storage and embeddings, retrieval-augmented generation); security across the AI life cycle; threat modeling with the OWASP Top 10 for LLM Applications, MITRE ATLAS and the MIT AI Risk Repository; guardrails, prompt firewalls and gateway limits (rate, token and input quotas); access control for models, data and agents; encryption, redaction, masking and data minimization; prompt and log monitoring; auditing for hallucinations, bias and accuracy; attacks (prompt injection, data and model poisoning, jailbreaking, model theft and inversion, AI supply chain, insecure output handling, excessive agency, overreliance) and the controls that answer them; AI-enabled security tools and automation; how attackers use AI; AI governance, roles and responsible AI principles; compliance frameworks (NIST AI RMF, EU AI Act, ISO and OECD standards).

Courses: the SecAI+ exam objectives are the outline. Read them in week 1 (the link is on the Resources page) and turn each objective into a question you can answer. The study guide and courses on the Resources page cover the rest. CompTIA also sells its own CertMaster learning products; they are optional.

Labs

1. The target: run a small open model with Ollama on the lab host. Ask the same question with a zero-shot, a one-shot and a multi-shot prompt, and write a system prompt that sets its role and limits. Then build a small question-answering script that retrieves from a folder of made-up runbooks (embeddings and a vector store). An AI assistant can draft the script; read every line before you run it.
2. Threat model: draw the data flow of that application (user, prompt, model, retrieved documents, answer) and mark where each threat lands, using the OWASP Top 10 for LLM Applications and MITRE ATLAS. Rate each threat by likelihood and impact and pick a control for the top five.
3. Break it: attack your own application with direct prompt injection, an instruction hidden inside a retrieved document, a jailbreak attempt and a request that makes it reveal its system prompt. Log each attempt, the answer and whether it worked.
4. Guardrails and limits: add a prompt template that keeps instructions apart from user text, an input check and an output check, and a small gateway script in front of the model with a rate limit and a token cap. Repeat lab 3 and record which attacks now fail and which still work.
5. Data and access: classify the made-up documents, then redact and mask the fake names and ID numbers before indexing. Run the application under its own least-privilege account, keep its key out of the scripts, and allow only the application host to reach the model's port.
6. Watch and audit: log every prompt and response, keep secrets out of the logs, and write three detection rules in your S11 log setup (injection phrases, unusually long prompts, a jump in token use). Check 20 answers against the source documents for hallucinations and write down how many were wrong.
7. AI in security work: give an AI assistant (Claude, from C1) a day of lab logs and ask for a summary and a draft incident ticket for one event. Check every claim against the logs and note what it got wrong. Then list how attackers use AI (deepfake voice and video, tailored phishing, automated reconnaissance) and which control from S6 or S11 answers each.
8. Govern it: write a one-page AI use policy for a small team: which AI tools are sanctioned, private versus public models, what data may go in, and who approves. Name the roles, and map three risks from your threat model to the NIST AI Risk Management Framework functions (Govern, Map, Measure, Manage).

Prove it: the threat model with its five controls, the attack log from labs 3 and 4 (what worked before and after the controls), the three detection rules, and the one-page policy. A partner reviews them and tries three attacks of their own on your application.

Certs: SecAI+ CY0-001 (securing AI systems 40%, AI-assisted security 24%, AI governance, risk and compliance 19%, basic AI concepts related to cybersecurity 17%).

Gate: SecAI+ (CY0-001) happens here, with one review week. Take CompTIA's free practice questions first, and review the securing AI systems domain twice, since it is the largest.