Basic Concept: Jailbreaking through prompt injection exploits the LLM ' s tendency to follow instructions embedded in user input, overriding its intended behavior. The model was manipulated to offer unauthorized discounts, demonstrating that its operational boundaries were not properly enforced. CompTIA SecAI+ Study Guide identifies guardrails as the primary defense against jailbreaking attacks.
Why D is Correct: Guardrails are robust, layered controls that enforce behavioral boundaries on LLM inputs and outputs. They can detect and block jailbreaking attempts, enforce business logic constraints such as preventing unauthorized discounts, validate outputs against policy rules before delivery, and prevent the model from operating outside its defined scope. Guardrails are specifically designed to make models more robust against prompt injection and jailbreaking.
Why A is Wrong: Bias filtering is designed to detect and remove biased, discriminatory, or offensive content from model outputs. It addresses content fairness issues but does not prevent jailbreaking attacks that manipulate the model into performing unauthorized actions.
Why B is Wrong: A system prompt sets the model ' s base instructions and persona, but the jailbreak attack already demonstrates that the current prompt can be overridden. Guardrails provide enforcement at a layer that is more resistant to prompt manipulation than the system prompt alone.
Why C is Wrong: Log monitoring detects jailbreaking attempts after they have already succeeded. It is a detective control that enables incident response but does not prevent the model from offering unauthorized discounts in the first place.

Basic Concept: Adding validation layers to software development processes improves security assurance by catching issues that human reviewers might miss. AI-assisted validation provides an automated, systematic review that complements human judgment. CompTIA SecAI+ Study Guide covers AI-assisted development security controls.
Why A is Correct: AI-assisted approval adds an intelligent automated review layer that works alongside existing human review. AI can systematically analyze code for security vulnerabilities, coding standard violations, dependency risks, and policy compliance with greater consistency and speed than manual review.
This creates a defense-in-depth validation approach where both AI and human reviewers must approve changes, catching issues that either layer might miss independently.
Why B is Wrong: A low-code plug-in provides simplified visual development tools that reduce the amount of manual code writing required. It is a development productivity tool, not a security validation layer for reviewing already-written code.
Why C is Wrong: Automated rollback is a deployment safety mechanism that reverts a deployment to the previous version when errors are detected after deployment. It is a recovery control, not a validation layer applied during the development review process.
Why D is Wrong: Regression testing verifies that new code changes have not broken existing functionality. It tests functional correctness, not security vulnerabilities, and does not add an AI-powered security validation capability to the existing human review process.

Basic Concept: Agentic AI systems that execute shell commands based on model-generated output are vulnerable to prompt injection attacks where malicious actors craft inputs that cause the agent to run unauthorized commands. Input validation using allowlists is a critical defense mechanism. CompTIA SecAI+ Study Guide covers agentic AI security controls.
Why A is Correct: Adding logic that validates shell commands against an approved allowlist before execution is the most direct and effective defense. This ensures only pre-approved, safe commands can be executed regardless of what the agentic system ' s model generates, preventing malicious command injection from reaching the operating system. This principle of allowlist-based input validation is a foundational secure agentic AI control.
Why B is Wrong: Deprecating and retraining the model is a lengthy process that addresses root cause training issues but does not provide immediate protection against ongoing injection attacks in the current deployed system.
Why C is Wrong: Modifying the application to ignore a specific tag merely removes one attack surface while leaving the system vulnerable to other injection vectors. It is not a comprehensive defense.
Why D is Wrong: Using only approved libraries controls which code libraries the agentic system can call, but does not validate or restrict the shell commands generated by the model at runtime based on arbitrary user input.

Basic Concept: Unstructured data such as free-form text, images, and audio does not have predefined labels or rigid schema. Training an AI model effectively on unstructured data requires techniques that can leverage patterns within the data itself or adapt a pre-trained model to new data types. CompTIA SecAI+ covers AI training methodologies under basic AI concepts.
Why B is Correct: Fine-tuning takes a pre-trained foundation model that has already learned rich representations from massive unstructured datasets and further trains it on a specific, potentially smaller unstructured dataset. This adapts the model to a new domain, task, or data type without requiring labeled data for every training example. Fine-tuning is the most practical and effective approach for working with unstructured data in modern AI development.
Why A is Wrong: Statistical learning typically refers to classical machine learning approaches that often assume structured, numerical data with defined features. These methods generally struggle with high- dimensional unstructured data without significant preprocessing.
Why C is Wrong: Supervised learning requires labeled training data where each example has an associated correct output label. Applying supervised learning to unstructured data requires extensive manual labeling, which is the opposite of working with raw unstructured data.
Why D is Wrong: Reinforcement learning trains models through reward signals based on actions taken in an environment. It is designed for sequential decision-making tasks and is not the standard approach for learning representations from unstructured data at scale.

RAG augments an LLM with external, up-to-date information retrieval, improving its factual accuracy and contextual scope-capabilities that directly strengthen the model's effectiveness for reconnaissance.