AI-Assisted Security Operations — CompTIA SecAI+ practice questions

Prediction drift refers to:

1. Changes in the distribution of the incoming input features
2. Changes in the relationship between inputs and true labels
3. Changes in the distribution of model outputs over time
4. Changes to the optimizer state carried between training runs

Answer: C — Changes in the distribution of model outputs over time

C) Correct — prediction drift is measured on the output side: the mix of scores or classes the model emits shifts, which is often the first signal available when ground truth is delayed. A) Incorrect — that describes data drift on the input distribution. B) Incorrect — a changed input-output relationship is concept drift. D) Incorrect — optimizer state belongs to training and is not monitored in production.

A SOC lead wants to counter over-reliance on an AI triage assistant without abandoning it. Which measure is MOST likely to work?

1. Display the assistant's verdict more prominently in the case view
2. Require analysts to record the evidence supporting their disposition
3. Remove the assistant's confidence score from the analyst interface
4. Rotate analysts through the triage queue on a shorter shift cycle

Answer: B — Require analysts to record the evidence supporting their disposition

B) Correct — requiring the analyst to articulate independent evidence forces engagement with the underlying data and makes uncritical agreement visible in the case record. A) increases the anchoring effect and would deepen the reliance the lead is trying to reduce. C) hides useful calibration information without changing the incentive to defer, and analysts may then trust the verdict even more blindly. D) addresses fatigue rather than deference, and a fresh analyst can rubber-stamp just as readily.

An analyst pastes the full body of a reported phishing email, including the recipient's name and internal project references, into an external AI assistant for analysis. The MOST significant concern is:

1. the assistant may produce a verdict that disagrees with the mail gateway
2. internal data has been disclosed to a third party outside the sanctioned path
3. the assistant cannot parse email headers as accurately as the mail gateway itself can
4. the analysis will take longer than running the message through a sandbox

Answer: B — internal data has been disclosed to a third party outside the sanctioned path

B) Correct — the act of pasting sends organisational content to an external processor, which is a data handling and acceptable-use violation regardless of how good the resulting analysis is. A) describes an ordinary disagreement between tools that triage exists to resolve. C) may be true in some cases but is a capability limitation, not the significant risk introduced by the action. D) frames the issue as efficiency, which is beside the point when confidential content has already left the boundary.

Six months after deployment, an alert-scoring model assigns systematically lower risk to activity from one regional office than to identical activity elsewhere. Which term BEST describes what monitoring should be watching for here?

1. Adversarial example generation against the scoring model
2. Model inversion against the underlying training data
3. Bias drift in the model's behaviour across subgroups
4. Overfitting of the model to its original validation split

Answer: C — Bias drift in the model's behaviour across subgroups

C) Correct — differential treatment of comparable activity across a population attribute, emerging over time, is bias drift, and monitoring must slice performance by subgroup to see it. A) describes crafted inputs designed to evade a model, which would not manifest as a consistent regional skew. B) concerns recovering training data from a model and has nothing to do with scoring disparity. D) is a training-time generalisation failure that would show at deployment rather than appearing gradually months later.

A deployed AI assistant that answers analyst questions about internal policy begins citing a control document that does not exist. Which monitoring control would MOST reliably have caught this?

1. Automatically checking cited sources against the document repository
2. Reviewing the volume of analyst queries submitted per shift
3. Tracking the assistant's average response latency over each release
4. Surveying analysts quarterly on their satisfaction with the assistant

Answer: A — Automatically checking cited sources against the document repository

A) Correct — a fabricated citation is detectable mechanically by resolving every reference against the authoritative store, which turns a subjective quality problem into a measurable check. C) monitors performance and would be unaffected by a confidently fabricated answer. B) measures usage, which may even rise while the assistant is producing wrong content. D) is a lagging, subjective signal that depends on an analyst happening to notice and remember.

A team is generating response playbooks with a language model. Which practice MOST reduces the risk of an unsafe playbook reaching production?

1. Requiring peer review and a dry run in a test tenant before enablement
2. Restricting generation to playbooks under a fixed number of total steps
3. Instructing the model to include a rollback note at the end of each play
4. Generating several variants and enabling the one with the fewest actions

Answer: A — Requiring peer review and a dry run in a test tenant before enablement

A) Correct — the danger is destructive or incorrect automated action, and human review plus execution against a non-production tenant is what surfaces that before real assets are touched. B) uses length as a proxy for safety, but a short playbook that isolates the wrong subnet is far more damaging than a long careful one. C) is helpful documentation yet does nothing to verify that the steps themselves are correct or authorised. D) optimises for minimal action rather than correctness, and the fewest-action variant may simply omit necessary containment.

Analysts have begun accepting AI-generated verdicts without checking the underlying evidence, and a missed intrusion is traced to this habit. This pattern is BEST described as:

1. automation bias, an over-reliance on the system's recommendation
2. alert fatigue caused by excessive detection rule volume
3. confirmation bias in the construction of the hunt hypothesis
4. model drift arising from a change in the input distribution

Answer: A — automation bias, an over-reliance on the system's recommendation

A) Correct — automation bias is the tendency to defer to an automated recommendation and to stop seeking disconfirming evidence, which is exactly the described behaviour. B) is a related but distinct condition in which volume dulls attention; here the problem is trust in a verdict, not the number of alerts. C) concerns seeking evidence that supports a prior belief during investigation, not deference to a machine. D) describes a change in the model's operating conditions rather than a change in human behaviour.

An organisation is evaluating a third-party AI security tool before deployment. Which evaluation step provides the STRONGEST evidence that it will perform in this environment?

1. Reviewing the vendor's published benchmark results and case studies
2. Confirming the vendor holds current information security certifications
3. Testing the tool against the organisation's own labelled historical data
4. Comparing the tool's feature list against competing offerings in the market

Answer: C — Testing the tool against the organisation's own labelled historical data

C) Correct — performance is environment-specific, so replaying the organisation's own labelled history measures the only thing that matters: how the tool would have behaved on this network. A) reports results on someone else's data, tuned for the vendor's own presentation. B) attests to the vendor's security posture, which is necessary for supply-chain assurance but says nothing about detection quality. D) compares stated capability rather than demonstrated performance, and feature parity does not imply equivalent accuracy.

A detection team wants to cut false positives on a noisy rule using a supervised classifier trained on historical analyst dispositions. Which risk is MOST important to plan for before the classifier is trusted to auto-close alerts?

1. The classifier will need substantially more storage than the rule it supports
2. Supervised methods are unable to output a confidence value for each alert
3. The classifier cannot be retrained once it is deployed into the pipeline
4. Past mislabelled closures will be learned and reproduced as confident suppressions

Answer: D — Past mislabelled closures will be learned and reproduced as confident suppressions

D) Correct — the labels are analyst dispositions, so any historical habit of closing a class of true positives becomes ground truth and the model industrialises that error at machine speed. A) is a true-but-irrelevant operational concern that has no bearing on whether auto-closure is safe. C) is simply false; retraining on newer dispositions is the normal maintenance cycle for such models. B) is also false — confidence scores are exactly what an auto-close threshold is built on.

An LLM that generates plausible but factually incorrect output is exhibiting:

1. Bias: outputs systematically skewed against a group
2. Hallucination: fluent output unsupported by facts
3. Drift: input distribution shifts away from training
4. Overfitting: memorized training data, poor on new

Answer: B — Hallucination: fluent output unsupported by facts

B) Correct — hallucination is confident, fluent generation that is not grounded in fact; a headline GenAI risk. A) Bias is a systematic skew across groups, not fabrication. C) Drift is a change in the data distribution over time. D) Overfitting is memorization of training data, a different failure.

Distinguish human-in-the-loop from human-on-the-loop supervision.

Human-in-the-loop requires human approval before each action executes. Human-on-the-loop lets the system act autonomously while a human monitors and can intervene or reverse. The choice follows reversibility and blast radius: irreversible high-impact actions demand in-the-loop.

Distinguish signature-based from anomaly-based detection.

Signature-based detection matches known indicators or patterns: high precision, low false-positive rate, blind to novel attacks. Anomaly-based detection models normal behaviour and flags deviation: capable of catching unknown attacks, but noisier and dependent on a clean, representative baseline.

What should be evaluated when assessing an AI security tool for purchase?

Measured precision and recall on your own data rather than the vendor's • explainability of each verdict • false-positive rate at your alert volume • how the model is updated and whether your data trains it • data residency and retention • integration and rollback story • degradation behaviour when the model is unavailable.

Data drift vs Concept drift?

Data drift: feature distribution changes (e.g., new user demographics). Concept drift: relationship between features and label changes.