CompTIA SecAI + CY0-001 (V1) Exam Practice Test 2

Hidden instructions embedded within content the model processes, rather than typed directly by the user, that the model then follows is an indirect form of prompt injection. Jailbreaking typically involves a user directly crafting input to bypass safety guardrails, model theft steals the model itself, membership inference determines whether specific data was used in training, and excessive agency concerns an AI system being granted more autonomous capability than appropriate rather than describing this specific injection technique.

Data sovereignty concerns legal requirements that data be stored and processed within specific geographic or jurisdictional boundaries, exactly as described with EU customer data. Third-party compliance evaluations assess vendor controls, sanctioned versus unsanctioned tools concerns internal approval status, private versus public models concerns model deployment architecture, and NIST AIRMF is a specific voluntary risk framework rather than a geographic data residency requirement.

Reproducing licensed proprietary code creates a direct intellectual property risk, since using such output commercially could infringe on someone else's rights. Reputational loss and accidental data leakage are different potential consequences of AI risk generally, and bias introduction and autonomous systems risk describe unrelated risk categories.

Requiring an independently audited report of a vendor's controls before contracting is a third-party compliance evaluation. Data sovereignty concerns where data is legally allowed to reside, sanctioned versus unsanctioned tools concerns internal approval status, EU AI Act compliance is a specific regional legal requirement, and NIST AIRMF alignment is a specific voluntary framework rather than the general practice of vendor compliance review.

Evaluating whether a model's decisions can be clearly explained, rather than functioning as a black box, is explainability. Differential privacy concerns protecting individual data through statistical noise, consistency concerns stable behavior across similar inputs, fairness concerns equitable treatment across groups, and awareness training concerns educating people rather than the model's own interpretability.

An AI auditor specifically reviews AI systems against requirements, verifying documentation, testing, and controls before and after deployment. An AI risk analyst assesses and prioritizes risk rather than performing formal compliance review, a platform engineer builds underlying infrastructure, a data engineer builds data pipelines, and an AI architect designs system architecture rather than auditing it.

An autonomous component that independently investigates, gathers evidence, and takes multi-step action without a human initiating each step is an AI agent. Low-code scripting describes a development approach rather than autonomous action, document synthesis condenses existing content, automated deployment/rollback concerns pushing or reverting changes, and unit testing validates individual code components.

Automatically extracting key findings from reports and drafting a plain-language summary is document synthesis and summarization. AI agents describes autonomous multi-step actors more broadly, change management concerns approving and deploying changes, incident response ticket management handles case tracking, and model testing validates AI model behavior rather than summarizing external reports.

Deliberately fabricated content spread with intent to deceive is disinformation, distinguished from misinformation, which is false content spread without necessarily intending to deceive. Impersonation involves posing as a specific real person or entity, reconnaissance gathers target information, and social engineering manipulates people directly rather than describing this content-fabrication campaign.

Rapidly cross-referencing disparate data sources to automatically build detailed target profiles is automated data correlation. Obfuscation hides malicious content from detection, a honeypot is a defensive decoy, adversarial networks refers to GAN-style competing model training, and distributed denial of service targets availability rather than building target profiles.

Flagging unusual patterns that deviate from a learned baseline, without matching a known signature, is anomaly detection. Signature matching relies on known patterns, translation converts between languages, summarization condenses content, and fraud detection specifically targets financial or transactional abuse rather than describing general network traffic anomalies.

An AI assistant integrated directly into a terminal environment is a CLI plug-in. An IDE plug-in integrates into a code editor rather than a terminal, a browser plug-in integrates into web browsing, a chatbot is typically a standalone conversational interface, and an MCP server exposes tool access to a model rather than being the terminal-integrated assistant itself.

Granting an AI agent autonomous capability far beyond what its actual intended function requires is excessive agency. Overreliance describes humans trusting AI output too much rather than the agent having too much capability, model skewing gradually biases model behavior, insecure plug-in design concerns flaws in how a plug-in is built rather than how much capability it was granted, and an output integrity attack tampers with output content rather than describing over-permissioning.

Federated learning trains a shared model across many decentralized devices or organizations without raw data ever leaving its original location, exactly matching this constraint. Supervised and unsupervised learning describe how labels are used during training rather than where data resides, reinforcement learning trains through reward signals, and transfer learning reuses a pretrained model on a new task rather than describing decentralized training.

Confidently worded but entirely fabricated details, such as referencing product features that don't exist, is the definition of hallucinations. Bias and fairness concerns disparate treatment across groups, access concerns who can use the system, and accuracy and accountability are broader or related but less precisely descriptive of this specific fabrication pattern than hallucinations.

Automatically stripping sensitive data like credit card or social security numbers from logs before long-term storage is log sanitization. Log monitoring observes log activity for issues, log protection secures logs from unauthorized access or tampering, prompt monitoring observes live query and response content, and rate monitoring tracks request volume rather than removing sensitive content from stored logs.

Breaking down spend across categories like storage, prompts, and processing is specifically AI cost monitoring. Rate monitoring tracks request volume rather than dollar cost, log monitoring tracks log activity broadly, response confidence level tracks model certainty, and prompt monitoring observes query and response content rather than billing breakdowns.

Removing direct identifiers so records can no longer be tied to specific individuals at all is data anonymization, while partially obscuring a value like a credit card number while retaining some visible portion is data masking. Classification labels tag data by sensitivity level rather than altering values, data minimization limits how much data is collected in the first place, and data redaction typically removes content entirely rather than partially obscuring it.

This scenario distinguishes model access, which governs the ability to query the deployed model, from data access, which governs the ability to reach the underlying training data, granting one without the other. Agent access concerns autonomous components rather than a human contractor, and network access and endpoint access describe connectivity layers rather than this specific query-versus-data-access distinction.

Restricting which input or output formats, such as text versus image or audio, are permitted is a modality limit. Rate limits cap request frequency, token limits cap content volume per call, input quotas cap the size or quantity of a given input type, and prompt firewalls filter malicious prompt content rather than restrict which data formats are accepted at all.

Automatically rewriting user input into a standardized, pre-approved structure before it reaches the model is the use of prompt templates, a form of model guardrail. Model evaluation assesses overall model performance, rate limits cap request volume, endpoint access controls govern who can reach the API, and input quotas cap the size or quantity of input rather than restructuring its format.

The OWASP LLM Top 10 specifically catalogs the most critical vulnerability categories unique to LLM-based applications, such as prompt injection. MITRE ATLAS catalogs adversary tactics and techniques rather than a top-vulnerability list, the MIT AI Risk Repository catalogs AI risks broadly, the CVE AI Working Group tracks formally disclosed vulnerabilities, and NIST AIRMF is a risk management framework rather than a vulnerability list.

Validation specifically confirms the model performs as expected once moved into its real, live operating context, distinct from evaluation, which happens earlier against test data. Deployment is the act of releasing the model itself, monitoring and maintenance is the ongoing tracking that follows validation, and data preparation happens much earlier in the life cycle.

Checking for and correcting missing values, duplicates, and formatting inconsistencies is data cleansing. Data lineage tracks a dataset's history, data provenance tracks its origin and ownership, data balancing adjusts class distribution, and data augmentation creates synthetic variations rather than fixing quality issues in existing records.

Removing weights that contribute little to the model's output is pruning, while converting weight precision to a lower bit-width to shrink memory footprint is quantization. Epoch adjustment concerns how many passes through the training data occur, federated learning concerns decentralized training location, and data augmentation creates synthetic training examples rather than modifying the trained model's own weights.