CAICSO study plan

Map the six AI governance control families (policy, roles, inventory, third-party, documentation, audit) to concrete artifacts a candidate could be asked to produce

Draft a one-page AI acceptable use policy covering prohibited uses, human oversight, and disclosure of AI-generated content

Build a compliance crosswalk table that aligns at least five obligations (internal policy, contractual, sector regulator, privacy law, export control) to evidence types

Explain how to assign risk ownership and accountability across business, security, legal, and data science roles using a RACI example

Apply a risk-tiering method to classify five hypothetical AI use cases into low, medium, high, and prohibited tiers with stated criteria

Construct a risk register with likelihood, impact, inherent risk, control, and residual risk columns for at least eight AI-specific risks

Perform a threat-modeling pass on a retrieval-augmented chatbot using STRIDE or an equivalent structured method

Identify which assessment outputs feed governance reporting versus engineering backlog, and justify the split

Diagram a defense-in-depth reference architecture for an LLM application spanning identity, input handling, model access, output filtering, and logging

Compare prompt injection, indirect injection via retrieved content, jailbreaking, and data poisoning by attacker goal, precondition, and primary mitigation

Select and justify controls for model access (gateway, rate limiting, key management) and for tool or agent permissions (least privilege, allowlists)

Evaluate the residual risk of an agentic workflow with write access to internal systems and propose three compensating controls

Trace data flows for training, fine-tuning, retrieval, and inference, and mark where personal or sensitive data enters and persists

Design a data minimization and retention approach for embeddings, vector stores, prompts, and logs, including deletion propagation

Sequence the security gates in an MLOps pipeline from data ingestion through deployment, with entry and exit criteria at each gate

Specify integrity controls for model artifacts and dependencies, including provenance, signing, and registry access control

Define AI-specific detection signals such as anomalous prompt patterns, extraction attempts, output policy violations, and drift indicators

Draft an incident classification scheme that distinguishes model misuse, data compromise, model compromise, and availability events

Write containment and recovery steps for a model extraction incident and for a poisoned-model discovery, including rollback decisions

Identify the evidence needed for post-incident review and regulatory notification timelines

Build a red-team test plan for an LLM feature with at least twelve test cases mapped to threat categories and expected defenses

Design evaluation metrics for robustness, output safety, and over-refusal, and state pass thresholds

Assess a third-party model or API provider using a structured questionnaire covering data use, subprocessors, model change notice, and incident terms

Produce a control gap analysis comparing current-state controls to a target framework and rank gaps by risk

Translate technical AI risk findings into a two-slide executive briefing with risk framing, decision requested, and cost of inaction

Prepare responses to likely stakeholder objections about AI security investment from finance, legal, and product audiences

Define three program metrics that demonstrate AI security program maturity and explain how each is measured

Map exam-style governance scenarios to the correct control domain and justify the classification

Complete a timed full-length practice assessment and score each domain separately

Re-study the two weakest domains using the artifacts produced in earlier weeks

Rewrite condensed one-page summaries for all eight domains from memory, then correct against source material

Review every missed question to identify whether the error was knowledge, misreading, or reasoning