CompTIA SecAI + CY0-001 (V1) Exam Practice Test 4

Revealing confidential internal information like pricing formulas through cleverly worded questions is sensitive information disclosure. Model theft copies the model itself, model denial of service targets availability, insecure plug-in design concerns architectural flaws in integrated plug-ins, and membership inference determines whether specific data was in the training set rather than describing general leakage of confidential business content through conversation.

Deciding that certain data may only be processed by a privately hosted model rather than a third-party public API is the private versus public models policy concept. Sanctioned versus unsanctioned tools concerns which tools are approved for use generally, data sovereignty concerns geographic data residency, third-party compliance evaluations assess vendor controls, and sensitive data governance is the broader category this specific hosting decision falls under rather than the precise concept being illustrated.

The NIST AI Risk Management Framework is a voluntary US government framework providing structured risk guidance without binding legal force. The EU AI Act is a binding legal regulation, ISO AI standards provide international technical guidance through a different standards body, OECD standards offer broader international policy principles, and GDPR governs personal data protection generally rather than AI risk management specifically.

Users trusting and acting on AI output without independent verification, even when it is wrong, is overreliance. Excessive agency concerns the AI system itself being granted too much autonomous capability rather than human trust behavior, model skewing concerns gradual bias introduction, shadow AI concerns unauthorized tool adoption, and autonomous systems risk concerns systems acting without human review rather than human trust in a system's advice.

Training employees to appropriately question AI output rather than blindly trust it is awareness training, a responsible AI principle focused on human readiness. Differential privacy is a technical data protection method, consistency concerns stable model behavior, explainability concerns whether decisions can be understood, and inclusiveness concerns fair treatment across diverse user groups rather than employee education specifically.

An AI architect designs the overall technical architecture connecting models, data pipelines, and infrastructure. An MLOps engineer builds and maintains the deployment pipelines specifically, a data engineer builds general data infrastructure, an AI governance engineer focuses on policy and compliance structures, and a platform engineer maintains underlying compute platforms rather than the overall AI system architecture.

Building a workflow entirely through a drag-and-drop interface without writing any traditional code is a no-code approach. Low-code still involves some manual code, even if minimal, CI/CD concerns automated build and deployment pipelines, and unit testing and model testing validate code or model behavior rather than describing how the workflow itself was built.

Automatically opening, categorizing, and routing security tickets based on content is incident response ticket management. AI agents describes autonomous multi-step actors more broadly, document synthesis and summarization condenses existing content, change management concerns approving and deploying changes, and model testing validates AI model behavior rather than ticket triage specifically.

Automatically generating many functionally equivalent but structurally different malware variants to evade signature detection is a form of automated attack generation. Adversarial networks refers specifically to GAN-style competing model training, reconnaissance gathers target information, social engineering manipulates people directly, and honeypot evasion concerns avoiding decoy systems rather than generating malware variants at scale.

A synthetically generated voice convincingly impersonating a real executive is deepfake impersonation. Reconnaissance gathers target information rather than generating impersonation content, obfuscation hides malicious content from detection, automated data correlation links data points together, and a honeypot is a defensive decoy rather than an impersonation technique.

Flagging new transactions based on learned characteristics of previously confirmed fraud cases is fraud detection specifically. Anomaly detection flags general deviations from a baseline without necessarily being fraud-specific, signature matching relies on matching known exact patterns, and translation and summarization concern language conversion and content condensation rather than transactional fraud analysis.

An always-available, general-purpose conversational assistant that analysts consult throughout the day, rather than being embedded in one specific application, is a personal assistant. IDE, browser, and CLI plug-ins are embedded within specific applications or environments, and an MCP server exposes structured tool access to a model rather than serving as the conversational assistant itself.

A transfer learning attack exploits vulnerabilities in a shared base model that propagate to every downstream task built on it, an output integrity attack tampers with generated output after the fact, insecure output handling fails to validate or sanitize model output before downstream use, and membership inference determines whether a specific record was part of the training set, each representing a distinct point in the AI attack surface.

Discovering natural groupings in unlabeled data is unsupervised learning, while training on explicitly labeled examples is supervised learning. Reinforcement learning trains through reward signals rather than labels, federated learning concerns decentralized training location, and model validation checks performance rather than describing either of these two training styles.

Crafting input specifically to bypass safety filters and misclassify harmful content as acceptable is circumventing AI guardrails, while gradually introducing bias into a model's behavior over time through subtle training data patterns is model skewing. Jailbreaking typically aims to produce prohibited content directly rather than misclassify moderation decisions, model inversion reconstructs training data, and data poisoning is a broader category that model skewing is a specific gradual form of, distinguished here by its slow, incremental nature.

Applying access controls and encryption to protect stored log data itself from breach is log protection. Log monitoring observes log activity for issues, log sanitization removes sensitive content from logs before storage, rate monitoring tracks request frequency, and prompt monitoring observes live query and response content rather than securing the stored log data itself.

Tracking requests per minute to detect a sudden spike indicating scraping or abuse is rate monitoring. AI cost monitoring tracks spend, response confidence level tracks the model's own certainty, log sanitization removes sensitive data from stored logs, and prompt monitoring observes query and response content rather than raw request frequency.

Tagging datasets by sensitivity level so downstream systems can apply appropriate handling automatically is data classification labeling. Anonymization removes identifiers, masking partially obscures values, redaction removes specific content entirely, and data minimization limits how much data is collected in the first place rather than labeling existing data by sensitivity.

Restricting what an autonomous agent is permitted to act on, such as limiting it to only the current conversation's data, is agent access. Model access concerns who may query the model itself, data access concerns broader data permissions not scoped to a specific autonomous actor, and network/API access and endpoint access controls concern network-layer reachability rather than the agent's scope of action.

Restricting which network zones or IP ranges can reach an endpoint at all, independent of downstream authentication, is an endpoint access control. Rate limits cap request frequency, token limits cap content volume, modality limits restrict input/output formats, and prompt firewalls filter prompt content rather than controlling network-level reachability.

A filtering layer specifically designed to detect and block known malicious prompt patterns before they reach the model is a prompt firewall. Rate limits cap request frequency, token limits cap content volume, modality limits restrict input/output formats, and input quotas cap data size or quantity rather than filtering for malicious content patterns specifically.

The MIT AI Risk Repository is a broad, curated inventory of AI risks spanning technical, ethical, and societal categories from an academic research institution. MITRE ATLAS focuses on adversarial tactics specifically, the OWASP LLM Top 10 focuses narrowly on LLM application vulnerabilities, the CVE AI Working Group tracks formally disclosed vulnerabilities, and NIST AIRMF is a US government risk management framework rather than an academic risk catalog.

Collecting real-world outcomes, including incorrect predictions, and using them to inform the next round of retraining is feedback and iteration. Data preparation and model development/selection happen earlier in the life cycle before deployment, business use case alignment happens at the very start, and model evaluation happens against a held-out test set rather than live production feedback specifically.

Creating additional synthetic training examples by transforming existing data, such as rotating or adjusting brightness of images, is data augmentation. Data balancing adjusts the proportion of classes represented, data cleansing fixes errors, data lineage tracks history, and data verification confirms accuracy rather than generating new synthetic examples.

Explicitly assigning a persona, such as a formal legal assistant, that shapes tone and behavior for the whole session is a system role. User prompts are the actual questions a person types, and one-shot, multi-shot, and zero-shot prompting describe how many examples are provided rather than assigning an overall persona.