CompTIA SecAI + CY0-001 (V1) Exam Practice Test 5

Determining whether a specific individual's record was part of the training dataset, based on subtle response differences, is membership inference. Model inversion reconstructs approximations of training data more broadly rather than confirming one specific record's presence, model theft copies the model itself, model skewing gradually biases behavior, and an output integrity attack tampers with output content rather than inferring training set membership.

Specific internal rules governing which sensitive data categories may be used as AI input, and under what conditions, is sensitive data governance. Sanctioned versus unsanctioned tools concerns which tools are approved for use, private versus public models concerns hosting architecture, data sovereignty concerns geographic data residency, and third-party compliance evaluations assess vendor controls rather than internal data-handling rules.

OECD standards are broad, internationally recognized policy principles for trustworthy AI developed through consensus among member governments, distinct from a binding regulation or technical standard. The EU AI Act is a binding legal regulation, ISO AI standards are technical standards from a standards body, NIST AIRMF is a specific US voluntary framework, and GDPR governs personal data protection generally rather than AI policy principles specifically.

Systematically ranking candidates lower for reasons unrelated to actual qualifications, due to training data patterns, is introduction of bias. Accidental data leakage concerns unintended data exposure, reputational loss is a potential downstream consequence rather than the core risk itself, autonomous systems risk concerns unsupervised system action, and shadow AI concerns unauthorized tool adoption rather than describing this specific unfair ranking pattern.

Evaluating whether a system performs comparably across different demographic groups rather than favoring the best-represented ones is fairness. Explainability concerns whether decisions can be understood, consistency concerns stable behavior across repeated identical inputs, differential privacy concerns statistical noise-based data protection, and awareness training concerns educating people rather than evaluating a system's demographic performance parity.

Building and maintaining the pipelines that collect, clean, and structure raw data is a data engineer's responsibility. A data scientist designs and analyzes models using that data, an AI architect designs overall system architecture, an MLOps engineer manages deployment pipelines specifically, and an AI risk analyst assesses risk rather than building data infrastructure.

Specifically validating a retrained model's accuracy and behavior before it replaces the current version is model testing, distinct from testing traditional application code. Unit testing validates individual code components, software composition analysis checks third-party dependencies, regression testing checks that code changes don't break existing functionality, and code scanning analyzes source code for flaws rather than validating a model's own behavioral performance.

Automatically scanning every new commit for known vulnerable patterns and outdated dependencies before merge is code scanning. Unit testing validates individual components' correctness, regression testing checks that changes don't break existing functionality, model testing validates AI model behavior specifically, and change management concerns the broader approval and deployment process rather than the specific act of scanning code for vulnerabilities.

Coordinating many compromised devices to maximize disruptive impact on a target while evading defenses is a distributed denial of service attack, here enhanced by AI-driven dynamic coordination. Reconnaissance gathers target information, social engineering manipulates people directly, obfuscation hides malicious content from detection, and a honeypot is a defensive decoy rather than an availability attack.

A decoy system populated with fabricated content, designed to waste an intruder's time and reveal their techniques, is a honeypot. Reconnaissance gathers information about a target rather than deploying a decoy, obfuscation hides malicious content from detection, automated data correlation links data points together, and payloads are the actual malicious content delivered in an attack rather than a defensive decoy.

Automatically flagging style violations, unused variables, and best-practice deviations in code is code quality and linting. Vulnerability analysis specifically targets security weaknesses rather than general style issues, pattern recognition identifies recurring structures more broadly, threat modeling maps potential attack paths conceptually, and fraud detection targets financial or transactional abuse rather than code style.

A standardized protocol exposing specific internal tools and data sources to a model in a structured way is an MCP server. Chatbots and personal assistants provide general conversational interfaces rather than structured tool exposure, and browser and IDE plug-ins integrate into specific applications rather than describing this standardized protocol-based integration approach.

A structural flaw in a plug-in's own construction that allows arbitrary command execution is insecure plug-in design. Excessive agency concerns a component being granted too much permitted capability rather than having an architectural flaw, model theft copies the model itself, sensitive information disclosure leaks confidential data through output, and model denial of service targets availability rather than describing a code-level flaw in a plug-in's implementation.

A compact model with far fewer parameters, designed to run efficiently on constrained devices, is a small language model. Large language models are far larger and typically require substantial compute, a GAN involves two competing networks, transformers describe an architecture rather than a size class, and statistical learning is a broader category rather than describing this compact, edge-optimized design specifically.

Detecting a surge of near-identical requests across many accounts in a short window is exactly what rate monitoring is designed to surface. AI cost monitoring tracks spend, log protection secures stored log data, response confidence level tracks the model's own certainty, and data classification labels sensitivity rather than detecting coordinated request volume patterns.

Checking whether outputs remain reliable and correct across a wide range of inputs, separate from demographic fairness, is an accuracy audit. Bias and fairness specifically examines disparate treatment across groups, hallucinations concerns fabricated content specifically, access concerns who can use the system, and accountability concerns ownership of outcomes rather than output correctness itself.

Deciding not to collect certain sensitive categories of data at all is data minimization, while de-identifying already-collected records so individuals cannot be traced back is data anonymization. Data masking partially obscures values rather than fully de-identifying them, classification labels tag data by sensitivity without altering collection or identity, and data redaction removes specific content from existing records rather than describing either not collecting data or de-identifying it.

Encrypting data while it moves across the network is encryption in transit, while encrypting it while sitting idle in storage is encryption at rest. Encryption in use protects data while actively being processed rather than moving or sitting idle, and data masking and anonymization alter data values rather than describing encryption states.

Capping the combined input and output token count for a single call is specifically a token limit, distinct from rate limits which cap request frequency over time. Input quotas cap raw data size or quantity rather than model-specific token units, modality limits restrict input/output formats, and endpoint access controls govern network-level reachability rather than per-call content volume.

A separate authentication layer specifically required to query the model itself, independent of broader session validity, is model access control. Data access concerns permissions over underlying datasets, agent access concerns autonomous components, and network/API access and endpoint access controls concern network-layer reachability rather than this model-specific authentication layer.

Constraining what categories of content a model is allowed to generate at all, such as never recommending medical dosages, is a model guardrail. Rate limits cap request frequency, endpoint access controls govern network reachability, input quotas cap data size or quantity, and modality limits restrict input/output formats rather than restricting output content categories.

A general conceptual approach for systematically identifying potential threats during design, before selecting any specific named tool, falls under the broader category of threat-modeling frameworks. The CVE AI Working Group and OWASP ML Security Top 10 are specific named resources rather than the general category, the MIT AI Risk Repository catalogs risks broadly, and NIST AIRMF is a specific risk management framework rather than describing threat modeling generically.

Formally confirming that a model's purpose supports a documented company priority before proceeding is business use case alignment, the earliest life cycle stage. Data collection and model development/selection happen after this alignment is confirmed, deployment happens much later, and monitoring and maintenance is an ongoing post-deployment activity rather than this initial justification step.

Embedding a marker specifically to identify content as AI-generated after the fact is watermarking. Data provenance tracks where data originated more broadly, data lineage tracks a dataset's full processing history, data verification confirms accuracy, and RAG retrieves external information to ground model responses rather than marking generated content as machine-made.

Providing three separate worked examples before the actual task is multi-shot prompting, while a reusable structured skeleton with placeholder fields is a template. Zero-shot prompting provides no examples at all, system roles assign a persona rather than a reusable input structure, and one-shot prompting provides exactly one example rather than three.