AI Governance, Risk and Compliance — CompTIA SecAI+ practice questions

AI governance documentation that should accompany a deployed model includes:

1. The source code repository and nothing further
2. Model card, intended use, metrics, monitoring
3. A single headline accuracy number for the model
4. Product marketing copy describing the capability

Answer: B — Model card, intended use, metrics, monitoring

B) Correct — deployment documentation should support audit: model card and datasheet, intended use, disaggregated metrics, limitations and an ongoing monitoring plan. A) Code alone gives an auditor no context. C) One aggregate number is not evidence of fitness. D) Marketing material is not a governance artifact.

Explainability technique SHAP differs from LIME because SHAP:

1. Model-specific explanations limited to tree-based learners
2. Game-theoretic attributions carrying consistency guarantees
3. An inability to handle tabular data containing mixed types
4. An algorithm identical to LIME, published under another name

Answer: B — Game-theoretic attributions carrying consistency guarantees

B) Correct — SHAP allocates credit using Shapley values from cooperative game theory, which gives its attributions additivity and consistency properties that a local surrogate fit does not guarantee. A) Incorrect — the kernel formulation is model-agnostic, and tree variants are an optimization rather than a limit. C) Incorrect — tabular data with mixed columns is a standard SHAP use case. D) Incorrect — LIME fits a local surrogate model and offers no such guarantees.

Explainability tool SHAP is BEST for:

1. Searching hyperparameter space for better settings
2. Per-feature attribution for a single prediction
3. Ingesting and validating raw data pre-training
4. Encrypting model artifacts stored in the registry

Answer: B — Per-feature attribution for a single prediction

B) Correct — SHAP assigns each feature a contribution to an individual prediction, giving post-hoc explanations for black-box models. A) Hyperparameter search is a tuning activity. C) Ingest validation is a data-integrity control. D) Artifact encryption protects the supply chain, not interpretability.

NIST AI Risk Management Framework functions are:

1. Govern, Map, Measure, Manage
2. Plan, Build, Deploy, Retire
3. Identify, Protect, Detect, Respond, Recover
4. Train, Test, Validate, Audit

Answer: A — Govern, Map, Measure, Manage

B) Generic SDLC. A) Correct — NIST AI RMF core functions are Govern, Map, Measure, Manage. C) That's NIST Cybersecurity Framework. D) Not the AI RMF.

A model card documents:

1. Server hardware and network topology of the host
2. Only the headline accuracy figure on the test set
3. Network addresses and firewall rules for the API
4. Intended use, training data, per-group results, limits

Answer: D — Intended use, training data, per-group results, limits

D) Correct — a model card records intended and out-of-scope use, training data, disaggregated performance, limitations and ethical considerations. A) Hardware belongs in infrastructure documentation. B) A single aggregate metric hides group-level failure. C) Network detail sits in the deployment runbook.

Measurement bias occurs when:

1. Proxy features systematically mismeasure the true construct
2. Every subgroup is equally represented within the training set
3. The network uses ReLU activations instead of sigmoid units
4. Hyperparameters are tuned on a held-out validation split

Answer: A — Proxy features systematically mismeasure the true construct

A) Correct — measurement bias comes from the stand-in variable: the recorded proxy diverges systematically from the thing you actually intended to capture. B) Incorrect — balanced representation is the absence of a different bias, not this one. C) Incorrect — activation choice is an architecture detail unrelated to construct validity. D) Incorrect — tuning on a validation split is standard practice, not a bias source.

A regulator asks why a credit-related model denied a specific application. The organisation can supply overall accuracy statistics but nothing about the individual decision. Which capability is MISSING?

1. Reproducibility of the model's training run from stored artefacts
2. Explainability sufficient to account for a particular model output
3. Availability monitoring for the service hosting the model endpoint
4. Anonymisation of the personal data used to train the model itself

Answer: B — Explainability sufficient to account for a particular model output

B) Correct — the request is for a decision-level justification, which requires per-instance explanation methods rather than aggregate performance figures. A) would let the organisation rebuild the model but still would not explain one applicant's outcome. C) concerns service uptime and has no bearing on justifying a decision. D) is a privacy control for training data, valuable in its own right but unrelated to accounting for an individual result.

A team plans to fine-tune a publicly available model obtained from an open repository and deploy it in a customer-facing service. Which supply-chain assurance step should come FIRST?

1. Verify the artefact's origin and integrity against a trusted publisher
2. Benchmark the model's accuracy on the intended customer workload
3. Configure rate limiting on the service endpoint before it goes live
4. Draft a model card describing the fine-tuned model's intended use

Answer: A — Verify the artefact's origin and integrity against a trusted publisher

A) Correct — provenance and integrity checking comes first because an artefact of unverified origin may carry a backdoor, and every later activity builds on that untrusted base. B) measures quality but a poisoned model can perform excellently on benchmarks while behaving maliciously on a trigger. C) is a sound production control that belongs later in deployment. D) is required documentation, but writing it about an unverified artefact simply records inherited assumptions.

A security architect is building an AI risk register. Which entry is MOST useful for driving treatment decisions?

1. A note that the organisation uses large language models in several places
2. A scored risk with owner, affected system, and agreed treatment and date
3. A list of all model versions currently deployed across the environment
4. A summary of the vendor's published responsible AI principles document

Answer: B — A scored risk with owner, affected system, and agreed treatment and date

B) Correct — a register drives action only when each entry carries an assessed level, a named owner and a committed treatment with a date, which is what turns awareness into work. A) records a fact with no risk statement, owner or action attached. C) is an inventory, a valuable input to the register but not itself a risk entry. D) documents supplier intent and belongs in due-diligence evidence rather than in the organisation's own risk assessment.

An organisation has documented AI risks and assigned metrics but has no defined owner, policy, or escalation path for AI risk decisions. Which NIST AI RMF function is MOST clearly underdeveloped?

1. Map
2. Measure
3. Govern
4. Manage

Answer: C — Govern

C) Correct — Govern is the cross-cutting function that establishes policies, roles, accountability and escalation, and its absence is exactly what the scenario describes. A) Map appears to be working, since the organisation has identified and contextualised its risks. B) Measure also appears functional given that metrics have been assigned to those risks. D) Manage would be the next weakness to appear, but it cannot be sustained without the ownership and authority that Govern supplies.

What does an AI management system standard require an organisation to establish?

A documented management system for AI: policy and objectives, defined roles and accountability, AI risk assessment and impact assessment processes, controls over the AI lifecycle, supplier and data governance, competence and awareness, internal audit, and management review with continual improvement.

What obligations attach to a high-risk system under the EU AI Act?

Risk management across the lifecycle, data governance and quality controls, technical documentation and record-keeping, transparency to deployers, human oversight, and appropriate accuracy, robustness and cybersecurity, together with conformity assessment before placing the system on the market.

Distinguish transparency, explainability and interpretability.

Transparency is disclosing that AI is being used and how the system was built and evaluated. Explainability is giving a human-understandable reason for a particular output. Interpretability is the degree to which the model's internal mechanics can themselves be understood. A model can be transparent about its design and still be uninterpretable.

What is a model card, and what does it typically document?

A concise standardised disclosure document for a model: intended use and out-of-scope use, training and evaluation data at a high level, performance across relevant subgroups, known limitations and ethical considerations, and version and contact details. It supports transparency and informed adoption decisions.