CompTIA SecAI + CY0-001 (V1) Exam Practice Test 3

Overwhelming a model's processing capacity with expensive queries to make it unresponsive is model denial of service. Model theft copies the model itself, sensitive information disclosure leaks confidential data through outputs, an output integrity attack tampers with output correctness, and insecure plug-in design concerns architectural flaws in integrated plug-ins rather than a volumetric availability attack.

Explicitly designating which AI tools are approved for use versus which are not is the sanctioned versus unsanctioned tools policy concept. Private versus public models concerns model deployment and hosting architecture, sensitive data governance concerns handling rules for sensitive data specifically, data sovereignty concerns geographic data residency, and third-party compliance evaluations assess vendor controls rather than internal tool approval status.

A fully autonomous system acting without human review, later difficult to explain, illustrates the autonomous systems risk category specifically. Accidental data leakage concerns unintended data exposure, reputational loss is a potential downstream consequence rather than the core risk category itself, shadow AI concerns unauthorized tool adoption, and bias introduction concerns unfair treatment rather than unexplainable autonomous decision-making.

Checking that a model behaves predictably and stably across repeated runs with the same input is consistency. Explainability concerns whether decisions can be understood, inclusiveness concerns fair treatment across diverse user groups, fairness concerns equitable outcomes across groups, and accountability concerns clear ownership of outcomes rather than output stability across repeated identical queries.

Formal, written expectations for evaluating, approving, and using AI tools are AI policies and procedures. An AI Center of Excellence is the coordinating body that often produces such policies rather than the documents themselves, an AI governance engineer and AI security architect are individual roles, and a platform engineer builds infrastructure rather than writing governance documentation.

An AI risk analyst specifically identifies, assesses, and prioritizes the risks a given AI system poses. An AI architect designs system architecture, an MLOps engineer builds deployment pipelines, a data engineer builds data infrastructure, and a platform engineer maintains underlying compute platforms rather than performing formal risk assessment.

Re-running previously passing tests to catch cases where new changes break existing functionality is regression testing. Unit testing validates individual components in isolation rather than checking for newly introduced breakage across the whole suite, software composition analysis checks third-party dependencies, model testing validates AI model behavior specifically, and code scanning analyzes source code for flaws rather than re-running a full prior test battery.

Automatically reverting to a previous stable version when error rates spike, without waiting for human intervention, is automated deployment/rollback. AI-assisted approvals concern pre-deployment sign-off decisions, code scanning and unit testing analyze code before deployment, and document synthesis condenses content rather than managing live deployment state.

Automatically identifying which specific ports and protocols are vulnerable before launching an attack is attack vector discovery, a component of automated attack generation. Payloads and malware generation refer to creating the actual malicious content or code, honeypot evasion concerns avoiding decoy systems, and reconnaissance is a broader information-gathering activity distinct from this specific vulnerability-discovery step.

AI-polished phishing content that better manipulates human targets by removing the usual red flags is an enhancement of social engineering. Reconnaissance gathers information about a target rather than crafting the deceptive message itself, obfuscation hides malicious code from technical detection, automated data correlation links data points together, and a honeypot is a defensive decoy rather than an attack technique.

Automatically attempting and adapting exploitation techniques to identify weaknesses is automated penetration testing. Vulnerability analysis typically identifies weaknesses without actively exploiting them, pattern recognition identifies recurring structures generally, threat modeling maps potential attack paths conceptually, and code quality and linting checks code style and correctness rather than actively attempting exploitation.

An AI feature integrated directly into a code editor is an IDE plug-in. A browser plug-in integrates into web browsing, a CLI plug-in integrates into a terminal, and a personal assistant and chatbot are typically standalone conversational interfaces rather than embedded directly within a code editing environment.

A hidden trigger embedded into model weights that activates malicious behavior only under a specific attacker-chosen pattern is a backdoor attack, while disguising a fully malicious model as a legitimate one to trick developers into downloading it is a trojan attack. Model skewing gradually biases behavior over time rather than embedding a discrete trigger, model inversion reconstructs training data, and membership inference determines whether specific data was used in training.

Learning through trial and error via reward or penalty signals from an environment is reinforcement learning. Supervised learning requires labeled examples, unsupervised learning finds patterns without labels, federated learning concerns decentralized training location, and model validation checks performance rather than describing this reward-driven training process itself.

Directly stealing and redeploying the actual trained model file is model theft. Model inversion reconstructs training data from model behavior rather than stealing the model file itself, membership inference determines whether specific data was in the training set, model skewing gradually biases behavior over time, and a transfer learning attack abuses a shared base model across tasks rather than directly copying a proprietary model.

Continuously watching both the queries sent and the responses generated is prompt monitoring, covering both the query and response side. Rate monitoring tracks request volume, AI cost monitoring tracks spend, log protection secures stored logs from tampering, and response confidence level tracks the model's own certainty rather than the content of queries and responses.

Checking whether only authorized roles could query the sensitive knowledge base is specifically an access audit. Response time under load is more accurately a performance or rate monitoring concern than one of the four named auditing focus areas of hallucinations, accuracy, bias/fairness, and access, so only the access half of this scenario maps cleanly to an auditing focus area; the performance-monitoring half falls under a different, operational monitoring category instead.

Completely removing specific sensitive content, such as a home address, rather than partially obscuring or replacing it, is data redaction. Data masking partially obscures a value while keeping some visible portion, anonymization removes identifiers while potentially retaining other content, classification labels tag data by sensitivity without altering it, and data minimization limits what is collected in the first place rather than removing existing content.

Restricting who can retrieve or export the underlying training dataset, separate from who can query the model itself, is data access control. Model access concerns querying the deployed model, agent access concerns autonomous components, and network/API access and endpoint access controls concern network-layer reachability rather than dataset export permissions specifically.

Measuring a model's accuracy, precision, and recall against a benchmark before production release is model evaluation. Model guardrails constrain runtime output behavior rather than benchmark performance, prompt firewalls filter malicious input, and rate limits and endpoint access controls govern request volume and API reachability rather than model performance assessment.

Capping the data size of a single request is an input quota, specifically limiting data size rather than request frequency. Rate limits cap requests per time window, token limits cap content volume in model-specific token units rather than raw data size, modality limits restrict input/output formats, and endpoint access controls govern who can reach the API at all.

The CVE AI Working Group specifically extends the formal, numbered CVE vulnerability tracking system to AI frameworks and libraries. MITRE ATLAS catalogs adversary tactics rather than individual disclosed vulnerabilities, the OWASP ML Security Top 10 lists common ML vulnerability categories rather than numbered disclosures, the MIT AI Risk Repository catalogs broad AI risks, and generic threat-modeling frameworks are not a specific vulnerability tracking catalog.

Converting text into numerical representations capturing semantic meaning is creating embeddings, and storing them in a database optimized for similarity search is vector storage, the two core components of a RAG pipeline. Fine-tuning and quantization modify a model's own weights, watermarking and data lineage concern provenance tracking, and data augmentation and balancing concern training data preparation rather than retrieval infrastructure.

Data with some organizational tags or metadata but without a rigid, fixed schema is semi-structured data. Structured data follows a strict, consistent schema like a database table, unstructured data has no organizational tags at all such as raw free text, and vector data and embedded data are not standard data type classifications in this objective's list.

The initial instruction defining persona, tone, and boundaries before any user interaction is a system prompt, while the actual question a customer types is a user prompt. One-shot prompting concerns providing a single example, templates concern standardized input structure, and system roles concern assigning a persona label rather than describing the full initial instruction set itself.