CompTIA SecAI + CY0-001 (V1) Exam Practice Test 1

Crafting user input that overrides the system's instructions to reveal confidential information is prompt injection, while subtly altering training data so the model misbehaves only under attacker-chosen conditions is data poisoning. Jailbreaking specifically aims to bypass safety guardrails to produce prohibited content rather than exfiltrate the system prompt, model inversion reconstructs training data from model outputs, and model theft steals the model itself rather than manipulating its training data.

The EU AI Act specifically classifies AI systems by risk level and imposes strict requirements on high-risk applications such as credit scoring. GDPR governs personal data protection broadly rather than AI-specific risk classification, NIST AIRMF is a voluntary US risk management framework, ISO AI standards provide international technical guidance, and OECD standards offer broader policy principles rather than binding EU risk-tiered legal requirements.

Unsanctioned AI tool use outside of IT visibility and approval processes is specifically shadow AI, a subset of the broader shadow IT problem. Reputational loss and accidental data leakage are potential consequences that could result from shadow AI use, but the situation itself, unauthorized unapproved tool adoption, is what defines shadow AI, and autonomous systems risk and bias introduction describe different, unrelated risk categories.

Adding calibrated statistical noise to protect individual records while preserving aggregate utility is the definition of differential privacy. Explainability concerns understanding how a model reaches its decisions, consistency concerns stable behavior across similar inputs, inclusiveness concerns fair treatment across diverse user groups, and accountability concerns clear ownership of outcomes rather than this specific noise-based privacy technique.

An MLOps engineer specifically builds and maintains the automated pipelines for deploying, monitoring, and retraining models in production. A data scientist focuses on model development and analysis, an AI architect designs overall system architecture, an AI security architect focuses specifically on securing AI systems, and a data engineer builds general data infrastructure rather than model-specific deployment pipelines.

A centralized internal group coordinating AI standards, tool evaluation, and initiatives across business units is an AI Center of Excellence. An AI governance engineer and AI risk analyst are individual roles rather than an organizational structure, AI policies and procedures are the documented rules the CoE typically helps produce rather than the coordinating body itself, and an MLOps engineer focuses on operationalizing machine learning pipelines.

Automatically inventorying third-party and open-source components and checking them against known vulnerabilities is software composition analysis. Unit and regression testing verify code correctness and that changes don't break existing functionality, model testing validates AI model behavior, and code scanning typically analyzes the organization's own source code for flaws rather than cataloging third-party dependencies.

Having an AI agent evaluate and automatically approve low-risk changes, while escalating higher-risk ones, is AI-assisted approvals. Automated deployment/rollback concerns actually pushing or reverting changes rather than approving them, code scanning and regression testing analyze code for issues, and model testing validates AI model behavior specifically rather than infrastructure change approval.

Using AI to automatically generate many payload variations designed to evade detection is automated attack generation. Adversarial networks specifically refers to GAN-style competing model training, social engineering manipulates people rather than generating payloads, reconnaissance gathers target information, and automated data correlation links data points rather than generating novel attack artifacts.

Convincingly synthesized audio and video used to impersonate a real person is deepfake impersonation, a specific category of AI-generated content attack. Reconnaissance gathers information about a target, obfuscation hides malicious content from detection, automated data correlation links disparate data points together, and a honeypot is a defensive decoy rather than an impersonation attack.

Identifying code with structural similarity to previously known vulnerable patterns is signature matching, applying known patterns to new content. Anomaly detection flags deviations from a learned baseline rather than matching known patterns, fraud detection targets financial or transactional abuse, translation converts between languages, and incident management coordinates response workflow rather than scanning code for known-bad patterns.

A browser-based extension that provides AI-driven analysis while browsing is a browser plug-in. An IDE plug-in integrates into a code editor, a CLI plug-in integrates into a command-line environment, a personal assistant provides broader conversational task support, and an MCP server exposes structured tool access to an AI model rather than running as a browser extension itself.

Model inversion reconstructs an approximation of training data by analyzing patterns in a model's outputs across many carefully chosen queries. Model theft copies the model's functionality or parameters rather than the training data, model skewing gradually biases a model's behavior, membership inference determines only whether a specific record was in the training set rather than reconstructing the data itself, and model denial of service overwhelms the model's availability.

Two competing networks training against each other to generate realistic synthetic content is the definition of a GAN, while a self-attention architecture converting sequences into contextualized representations, underlying most modern LLMs, is a transformer. Statistical learning and deep learning are broader categories, and reinforcement learning trains an agent through reward signals rather than adversarial competition or self-attention.

Checking whether a model treats comparably qualified candidates from different demographic groups similarly is specifically a bias and fairness audit. Hallucinations concern fabricated outputs, accuracy concerns correctness of outputs generally, access concerns who can use the system, and accountability is a broader responsible AI principle rather than this specific demographic-parity check.

Response confidence level tracks the model's own certainty scores over time, and a drop from baseline is exactly the kind of signal this monitoring concept is designed to catch. Rate monitoring tracks request volume, AI cost monitoring tracks spend, log protection secures log data from tampering, and prompt monitoring observes the content of queries and responses rather than the model's confidence scoring specifically.

Encrypting data while it sits in cloud storage is encryption at rest, encrypting it while it travels between systems is encryption in transit, and keeping it encrypted even while actively being processed in memory is encryption in use, a more advanced protection often implemented through confidential computing. Data masking and anonymization alter or obscure data values rather than describing encryption states across its lifecycle.

Network/API access controls govern which systems or callers can reach an endpoint at the network and API layer, independent of what the model is asked to do once reached. Model access concerns which authenticated users or roles may query a model's outputs, data access concerns what data the model can read or write, agent access concerns what autonomous components may act on, and model guardrails constrain output content rather than network-level reachability.

Agent access specifically governs what autonomous AI agents are permitted to do on their own, such as merging code without human review, which is exactly the concern here. Model access controls who can query the underlying model, data access controls what data the system can read or write, network/API access controls network-level reachability, and model guardrails constrain model output content rather than an agent's permitted actions.

Running known attack prompts against a guardrail before deployment to confirm it blocks malicious input without breaking legitimate use is guardrail testing and validation. Model evaluation assesses the underlying model's overall performance rather than a specific guardrail, prompt monitoring and rate monitoring are ongoing production activities, and log sanitization removes sensitive data from logs rather than testing a guardrail.

Rate limits cap how many requests can be made in a given time window, preventing a flood of automated requests, while token limits cap how much content a single call can consume or generate, preventing one crafted prompt from being excessively expensive. Prompt firewalls filter malicious prompt content rather than controlling volume, modality limits restrict input/output types like image or audio, and endpoint access controls govern who can reach the API at all.

MITRE ATLAS is explicitly modeled after ATT&CK's tactic-and-technique structure but focused on adversarial attacks against AI and machine learning systems. The OWASP LLM Top 10 lists the most critical LLM-specific vulnerabilities rather than a broad tactic catalog, the CVE AI Working Group tracks formally disclosed AI-related vulnerabilities, the MIT AI Risk Repository catalogs AI risks broadly, and NIST AIRMF is a risk management framework rather than an adversary tactic knowledge base.

Requiring a human to review and approve every AI-generated output before action is taken is human-in-the-loop, where a human is directly embedded in the operational decision path. Human oversight is a broader, less transaction-specific supervisory role, human validation checks outputs after the fact rather than gating every action, and feedback/iteration and model evaluation are later lifecycle activities rather than a real-time approval gate.

Data lineage tracks the full history of a dataset, including its sources, transformations, and processing steps, exactly as described. Data cleansing removes errors and inconsistencies, data balancing adjusts class distribution, data augmentation creates synthetic variations of existing data, and data verification confirms accuracy rather than tracing origin and transformation history.

Providing exactly one example before the actual task is one-shot prompting. Zero-shot prompting provides no examples at all, multi-shot prompting provides several examples, a system prompt sets overall behavior rather than providing task examples, and template prompting is not a standard named technique in this objective's list.