Basic Concept: LLM hallucinations occur when the model generates plausible-sounding but factually incorrect or fabricated information. For internal content management solutions where accuracy is critical, detecting and handling hallucinated responses before they are acted upon is essential. CompTIA SecAI+ Study Guide covers response validation as a mitigation for hallucination risks.
Why B is Correct: Response validation implements checks that verify the accuracy and relevance of LLM- generated responses before they are presented to users or acted upon. This can involve cross-referencing responses against authoritative internal data sources, using a secondary model to evaluate response accuracy, or implementing confidence scoring that flags low-confidence responses for human review. Response validation directly addresses the hallucination problem by catching inaccurate responses before they cause harm.
Why A is Wrong: Model training addresses hallucinations at the model level by providing more accurate training data or fine-tuning. While effective long-term, it requires significant time and resources and does not provide immediate protection against hallucinations in the currently deployed model.
Why C is Wrong: Access controls manage who can query the LLM and what resources they can access. They do not inspect or validate the accuracy of the model ' s responses, so they cannot mitigate hallucination risks.
Why D is Wrong: Integrity monitoring tracks whether data or systems have been tampered with or changed unexpectedly. It is relevant for detecting unauthorized modifications but does not validate whether LLM- generated content accurately reflects reality or internal authoritative data.

Basic Concept: AI systems can inadvertently reveal sensitive information such as PII, credentials, or internal data in their outputs when not properly controlled. Sensitive information disclosure is a critical OWASP LLM Top 10 risk. CompTIA SecAI+ Study Guide covers both vulnerability identification and appropriate data protection controls for AI outputs.
Why D is Correct: The scenario describes the AI system outputting sensitive information in its responses, which is a sensitive information disclosure vulnerability. The appropriate control is masking, which replaces sensitive data values such as credit card numbers, SSNs, or API keys with redacted or tokenized equivalents in the model ' s outputs before they are returned to users. This prevents the AI from disclosing sensitive data while still providing useful responses.
Why A is Wrong: Prompt injection involves crafting inputs to override model instructions. If the penetration test revealed sensitive information, the primary vulnerability is the disclosure of that sensitive data, not the injection mechanism itself. EDR monitors endpoint behavior, not AI output content.
Why B is Wrong: Model hallucinations produce fabricated information rather than disclosing real sensitive data. The described scenario involves actual sensitive information being revealed, not fictitious content generation.
Why C is Wrong: Jailbreaking circumvents safety restrictions but the primary harm demonstrated is sensitive data exposure. RBAC manages access permissions but does not prevent the model from including sensitive data in responses once access is granted.
Why E is Wrong: Role impersonation involves the AI pretending to be a different entity. This may be a secondary technique used by the penetration tester but the primary vulnerability described is the disclosure of actual sensitive information in the output.

Basic Concept: Assessing attack probability and impact for ML systems requires a resource specifically built to catalog real-world adversarial attacks against AI and ML systems, including documented techniques with associated impact information. CompTIA SecAI+ Exam Objectives identify MITRE ATLAS as the authoritative ML threat landscape resource.
Why B is Correct: MITRE ATLAS is specifically designed as a comprehensive knowledge base of adversarial tactics, techniques, and case studies targeting AI and ML systems. It catalogs real-world attacks with associated probability factors derived from actual incidents and provides impact assessments for various attack types including data poisoning, model evasion, model extraction, and inference attacks. This directly enables the probability and impact assessment the team requires.
Why A is Wrong: The CVE AI working group focuses on identifying and cataloging specific vulnerability instances in AI software components. While useful for vulnerability management, it does not provide the comprehensive threat landscape coverage with probability and impact assessments for ML-specific attack tactics that ATLAS provides.
Why C is Wrong: The MIT risk repository is an academic resource cataloging general AI-related risks. It is research-oriented and does not provide the practitioner-focused, operational attack taxonomy and case study library that MITRE ATLAS offers for ML threat modeling.
Why D is Wrong: OWASP provides application security guidance including the OWASP LLM Top 10. While valuable for LLM-specific risks, OWASP does not provide the comprehensive ML threat landscape coverage or the probability and impact data that MITRE ATLAS offers for assessing the full spectrum of ML attack scenarios.

Basic Concept: Different ML model architectures offer varying degrees of explainability. In cybersecurity, understanding why a model classified a log entry as malicious or benign is critical for analyst trust, investigation, and regulatory compliance. CompTIA SecAI+ covers model explainability under responsible AI and basic AI concepts.
Why C is Correct: Decision trees are inherently interpretable models that classify data through a series of transparent if-then rules. Every classification decision can be traced through the exact path of conditions that led to it, showing precisely which log entry features triggered the classification. Analysts can read and understand the decision path, making decision trees the gold standard for explainable ML classification in security applications where understanding the reason for a classification is as important as the classification itself.
Why A is Wrong: Large language models are complex transformer architectures with hundreds of billions of parameters. They function as black boxes - their internal decision-making processes are not human- interpretable, making them poor choices when explainability is the primary requirement.
Why B is Wrong: Neural networks are non-linear black box models. While they can achieve high classification accuracy, their multi-layer architecture makes it extremely difficult to explain why specific decisions were made in human-understandable terms.
Why D is Wrong: Generative adversarial networks are designed for generating synthetic data, not for classification tasks. They consist of competing generator and discriminator networks and are fundamentally unsuitable for log entry classification with explainability requirements.

Basic Concept: Threat modeling for any system, and especially for agentic AI systems with multiple interacting components, begins with understanding the system ' s architecture and where trust boundaries exist. Trust boundaries define where data and control flows cross between components with different trust levels, representing potential attack surfaces. CompTIA SecAI+ Study Guide aligns with STRIDE and MITRE ATLAS threat modeling methodologies.
Why B is Correct: Identifying trust boundaries between components is the foundational first step in threat modeling. Agentic systems often involve multiple components such as the orchestrator, tools, APIs, data sources, and external services with different trust levels. Understanding where these boundaries exist reveals where untrusted inputs cross into trusted components, enabling the architect to systematically identify threats at each boundary before proceeding to risk quantification and control application.
Why A is Wrong: Applying compensating controls based on exposure findings is the final step in threat modeling, occurring after threats have been identified and risks quantified. Controls cannot be appropriately designed without first understanding the system ' s trust boundaries and threat landscape.
Why C is Wrong: Calculating risk to resources based on data sensitivity is a risk assessment step that occurs after trust boundaries are mapped and potential threats are identified. Risk quantification requires knowing what threats exist at each boundary first.
Why D is Wrong: Scanning for OWASP Top 10 vulnerabilities is a technical vulnerability assessment activity. While valuable, it comes after the architectural analysis of trust boundaries and threat identification phases of threat modeling.