Free CompTIA SecAI+ Sample Questions

Correct answer: A — Model drift

• A. Model drift (specifically concept or data drift) refers to the degradation of a model's predictive power over time as the statistical properties of the data change. In cybersecurity, this typically occurs when adversaries evolve their tactics—such as using new obfuscation techniques—to bypass existing detection patterns that were established during the model's initial training phase.
• B. Overfitting occurs when a model learns the training data, including its noise and outliers, too closely, which prevents it from generalizing well to new, unseen data. While overfitting causes poor performance, it usually results in poor results immediately upon deployment; it does not explain why a previously 'highly accurate' model would degrade months later.
• C. Underfitting happens when a model is too simplistic to capture the underlying patterns in the data, leading to poor performance on both training and real-world data. This does not apply here because the model was initially highly accurate, indicating it had successfully captured the relevant patterns at the time of deployment.
• D. Tokenization is a fundamental Natural Language Processing (NLP) preprocessing step where text is broken down into smaller units, such as words, phrases, or symbols. While it is used in building phishing detection models, it is a data preparation technique and is not a cause for performance degradation over time.

Correct answer: B — Lack of explainability (Black box)

• A. High bias occurs when a model is overly simplified and fails to capture the underlying patterns in the data, leading to underfitting. The scenario describes a lack of transparency in reasoning, not a failure to learn from the data.
• B. Deep learning models are often described as 'black boxes' because their complex internal structures, involving many layers and millions of weight parameters, make it extremely difficult to interpret or explain the specific reasoning behind an output. This creates significant challenges for compliance, auditing, and trust in security environments.
• C. Deterministic output means that the model will produce the same result every time it is provided with the same input. While a model might be deterministic, this characteristic does not explain the difficulty in understanding the logic used to arrive at that result.
• D. Deep learning models are highly effective at processing both structured and unstructured data (such as network traffic, images, and text). This scenario focuses on the interpretability of the model's logic, not the format of the data it consumes.

Correct answer: A — Word embeddings

• A. Word embeddings represent words as dense numerical vectors in a multi-dimensional space where semantically similar words are positioned close together. This allows the AI to capture contextual relationships and identify that terms like 'exploit' and '0-day' share a similar meaning or context, even if they don't share a root word.
• B. Stemming is a text normalization technique that reduces words to their root form (e.g., 'exploiting' becomes 'exploit'). While it helps in grouping different forms of the same word, it does not encode semantic relationships through numerical vectors.
• C. Stop word removal is a preprocessing step that eliminates common words like 'the', 'is', and 'and' to reduce noise in the dataset. It does not involve mapping words into a vector space or establishing semantic relationships between terms.
• D. Optical Character Recognition (OCR) is a technology used to convert images of text into machine-readable text formats. It is a computer vision task and is not related to the semantic representation of word meanings within Natural Language Processing.

Correct answer: B — AI-driven polymorphic malware

• A. Data poisoning involves injecting malicious data into a machine learning model's training set to manipulate its future behavior or create backdoors. This does not match the behavior of malware mutating its own code on endpoints to bypass security software.
• B. Polymorphic malware changes its identifiable features (like its file hash and code structure) with each replication or infection to bypass signature-based detection systems. When enhanced by AI, this process becomes more sophisticated and efficient, allowing the malware to maintain its core logic and command-and-control (C2) communication while appearing unique to security tools.
• C. Automated exploit generation refers to using AI or automation to discover software vulnerabilities and craft functional exploit code. This scenario describes malware that has already been deployed and is mutating its code to evade detection, rather than the initial creation of an exploit.
• D. Membership inference is a privacy-focused attack where an adversary attempts to determine if a specific data record was part of a model's training dataset. It is irrelevant to the scenario of malware evasion and signature mutation.

Correct answer: B — Deepfake audio cloning

• A. Automated smishing involves the use of fraudulent SMS (text) messages to trick individuals. Since this scenario describes a voice-based phone call (vishing), smishing is not the correct term.
• B. Deepfake audio cloning uses AI and machine learning to replicate a specific person's voice, speech patterns, and cadence with high accuracy. This technology allows attackers to impersonate trusted individuals, such as executives, to perform social engineering attacks and authorize fraudulent financial transactions.
• C. Adversarial prompt injection involves manipulating the inputs or prompts given to an AI model (like an LLM) to bypass filters or force unintended outputs. It does not relate to the synthesis of human voices for social engineering.
• D. Polymorphic voice mutation is not a recognized or standard term in cybersecurity. While 'polymorphic' is used to describe malware that changes its code to evade detection, it is not used to describe the AI-driven synthesis of a person's voice.

Correct answer: A — AI-driven automated smishing leveraging real-time OSINT.

• A. Correct. This scenario illustrates the use of AI/automation to monitor Open Source Intelligence (OSINT) from social media feeds in real-time. By automatically identifying specific events (like conference attendance) and instantly generating tailored SMS messages, the attacker conducts highly personalized smishing (SMS phishing) attacks that are far more convincing than generic spam.
• B. Incorrect. Polymorphic malware refers to malicious software that changes its code or signature each time it replicates to evade detection. While malware might be delivered via the link, the scenario specifically describes the automation and targeting vector of the delivery mechanism itself.
• C. Incorrect. Adversarial evasion typically involves manipulating inputs to bypass machine learning classifiers or spam filters. While bypassing filters might be a sub-goal, the primary threat vector described is the automated scraping and context-aware messaging targeting a specific individual.
• D. Incorrect. Model extraction is an attack against an AI model where an adversary attempts to reconstruct or 'steal' the model's parameters or logic by analyzing its responses to specific queries. This scenario involves using AI for social engineering, not attacking a model's architecture.

Correct answer: B — Computer Vision

• A. Incorrect. Natural Language Processing (NLP) focuses on analyzing, understanding, and processing human language in text or speech formats (e.g., scanning emails or chat messages). While text-based filters often use NLP, they cannot interpret the visual pixel data within a screenshot.
• B. Correct. Computer Vision is the AI capability used to analyze, interpret, and understand visual content, such as digital images and videos. In this scenario, Computer Vision is required to scan screenshots and identify sensitive patterns (like credit card numbers) that traditional text-based filters would miss.
• C. Incorrect. Predictive Analytics uses historical data and statistical modeling to forecast future outcomes or trends, such as identifying likely fraud or system failures. It does not possess the image-processing capabilities needed to detect data exfiltration via screenshots.
• D. Incorrect. Generative AI is designed to create new, original content, such as text, images, or code. The objective here is the detection and classification of existing sensitive information within an image, not the generation of new content.

Correct answers: A, B — A high rate of false positives if the model is trained on poor or biased data; Susceptibility to adversarial attacks, such as data poisoning during the training phase

• A. Correct. Poor, insufficient, or biased training data often leads to inaccurate predictions. In security operations, this manifests as a high rate of false positives, which can cause alert fatigue and diminish trust in the AI system.
• B. Correct. AI systems are vulnerable to adversarial manipulation. Data poisoning involves an attacker corrupting the training data to influence the model's future behavior, potentially causing it to ignore specific malicious activities or misclassify threats.
• C. Incorrect. One of the primary advantages of AI is its superior ability to process and analyze massive volumes of telemetry and log data at a scale and speed that human analysts cannot match.
• D. Incorrect. AI is not infallible. While it can identify patterns associated with zero-day threats, it cannot guarantee perfect accuracy. AI models are probabilistic and will always have a margin of error.
• E. Incorrect. AI is intended to augment human capability, not replace it. Human oversight remains essential for complex investigations, strategic decision-making, and validating the context behind AI-generated alerts.

Correct answer: A — predictive threat modeling

• A. Predictive threat modeling uses AI to synthesize historical data, threat intelligence, and telemetry to anticipate and forecast future attacks, targets, or attack paths. This allows security teams to prioritize defenses and harden specific assets before an incident occurs.
• B. Reactive incident response occurs after a security event has already happened, focusing on containment, eradication, and recovery. It is the opposite of the proactive forecasting activity described in the scenario.
• C. Static signature analysis relies on identifying known patterns or files (signatures) to detect threats. It is a traditional detection-based method rather than a forecasting tool, and it lacks the AI-driven capability to analyze diverse datasets for future predictions.

Correct answer: B — Trusted Execution Environment (TEE)

• A. Homomorphic encryption allows computation on encrypted data without decrypting it first. While it protects data in use, it is typically far more computationally expensive than other methods and is not the standard choice for general-purpose cloud inference or protecting against an untrusted hypervisor via hardware isolation.
• B. A Trusted Execution Environment (TEE), often associated with Confidential Computing, provides a secure, hardware-isolated area within the processor (an enclave). This ensures that code and data are protected with respect to confidentiality and integrity, remaining encrypted in memory even from the host operating system or hypervisor.
• C. Secure Multi-Party Computation (SMPC) allows multiple parties to jointly compute a function over their inputs while keeping those inputs private. It is used for collaborative privacy but does not address the specific architectural requirement of securing an execution environment against an untrusted cloud infrastructure provider.
• D. A Hardware Security Module (HSM) is a physical device used to safeguard and manage digital keys and perform cryptographic operations. While it provides strong security for keys, it is not designed to host or execute full AI models or protect large amounts of data in memory during inference.

Correct answer: B — Implementing cryptographic model signing and hash verification

• A. Encrypting the model with AES-256 ensures confidentiality, but it does not provide a robust mechanism to verify integrity or authenticity. While it protects the model from being read, a tampered encrypted file could still be decrypted into a modified model if a separate integrity check is not implemented.
• B. Cryptographic model signing and hash verification are the industry standard for ensuring integrity and authenticity. By comparing the file's hash or digital signature against the one generated by the QA team, the DevOps engineer can verify that the model has not been modified or tampered with since its approval.
• C. A Trusted Execution Environment (TEE) provides hardware-level isolation and security for data and code during execution (runtime). While it protects the model while it is running, it does not inherently provide a mechanism to verify the integrity of the model file on disk prior to deployment.
• D. Differential privacy is a technique used to protect the privacy of individuals in the training data by adding noise to the model's outputs or weights. It does not address model file integrity or provide protection against unauthorized tampering of the model file.

Correct answer: A — Principle of Least Privilege

• A. Correct. The Principle of Least Privilege (PoLP) involves granting the minimum level of access and permissions necessary to perform a specific job function. By restricting the scientist's broad administrative access to only the specific S3 bucket required for their work, the administrator reduces the attack surface and potential for accidental or malicious damage.
• B. Incorrect. Separation of Duties (SoD) involves dividing tasks and responsibilities among different individuals to prevent fraud, errors, or abuse of power. While it is a critical security control, it does not specifically address the process of narrowing a single user's permissions to a specific resource.
• C. Incorrect. Defense in Depth is a strategy that uses multiple layers of security controls (such as firewalls, encryption, and monitoring) to protect an environment. While configuring IAM permissions is one layer of security, the specific act of limiting access to a single resource is governed by the Principle of Least Privilege.
• D. Incorrect. Security by Obscurity involves relying on the secrecy of design or implementation details to achieve security. This is not a recommended practice and does not address the requirement of formal access control or restricting user permissions.

Correct answer: A — Role-Based Access Control (RBAC) at the retrieval layer

• A. Correct. This is Role-Based Access Control (RBAC) applied at the retrieval layer, which is common in Retrieval-Augmented Generation (RAG) architectures. It ensures the system only fetches embeddings or records that the user is authorized to see, preventing unauthorized disclosure before the data is even processed by the model.
• B. Incorrect. Gradient masking is a technique used to improve adversarial robustness by hiding or obfuscating the model's gradients to prevent attackers from crafting adversarial examples. It does not address user-level permissions for document retrieval.
• C. Incorrect. Output obfuscation occurs at the presentation layer to hide or alter sensitive information after the model has generated a response. The scenario specifically describes filtering the input/retrieval source based on permissions, not the final output.
• D. Incorrect. Data sanitization at the training layer involves scrubbing or cleaning sensitive data before it is used to train a model. This scenario focuses on controlling access to existing stored documents during the inference phase.

Correct answer: A — Model watermarking

• A. Model watermarking involves embedding hidden identifiers or secret trigger-based behaviors into the model during training. This allows the owner to demonstrate provenance or ownership by providing the trigger phrase and verifying the specific, unique output (such as a copyright string) produced by the model.
• B. Output obfuscation refers to techniques that make the model's outputs less clear or harder to interpret for unauthorized users. While it can protect the utility of the output, it does not provide a verifiable mechanism for proving model provenance or original ownership.
• C. Gradient masking is a defense mechanism intended to make a model more robust against adversarial attacks by obscuring the gradients used by attackers to craft adversarial examples. It is used for robustness, not for establishing ownership or copyright attribution.
• D. Defensive distillation is a training method used to increase a model's robustness against adversarial examples by smoothing decision boundaries through training on 'soft' outputs from a teacher model. It does not involve secret triggers for ownership verification.

Correct answer: A — Randomized smoothing

• A. Randomized smoothing is a defense technique that provides provable robustness guarantees. It involves adding random noise (typically Gaussian) to an input at inference time and aggregating the predictions across many noisy samples (e.g., via a majority vote or averaging). This creates a 'smoothed' version of the classifier that is resistant to small L2-norm adversarial perturbations.
• B. Feature squeezing is a defense that reduces the complexity of input data—for example, by reducing color bit depth or applying spatial smoothing—to limit the degrees of freedom available to an adversary. While it helps detect adversarial examples, it does not involve the specific process of adding Gaussian noise and averaging predictions for formal robustness guarantees.
• C. Gradient masking (or gradient obfuscation) is a technique that attempts to hide the model's gradients to hinder optimization-based adversarial attacks. It is often considered an ineffective or 'brittle' defense because it can be bypassed by attacks that do not rely on direct gradients. It does not match the noise-sampling methodology described.

Correct answer: A — Sigstore / Cosign

• A. Sigstore and its tool Cosign are industry-standard tools specifically designed for signing and verifying container images and related artifacts. They provide cryptographic signatures that ensure the provenance and integrity of OCI images, making them essential for securing the software supply chain of containerized AI models.
• B. HashiCorp Vault is a tool used for managing secrets (such as API keys, passwords, and certificates), encryption, and secure storage. While it can be used for key management and can integrate with signing workflows, it is not the primary standard tool specifically used for the cryptographic signing of OCI images.
• C. Prometheus is an open-source monitoring and alerting toolkit used to collect metrics and observe the health of systems. It does not provide cryptographic signing or provenance verification for container images.
• D. Apache Kafka is a distributed event streaming platform used for building real-time data pipelines and streaming applications. It is unrelated to image signing or ensuring the integrity of containerized software artifacts.

Correct answer: B — A mantrap with biometric authentication

• A. CCTV cameras serve as a deterrent and a detective control for monitoring and forensic evidence collection. However, they are passive and do not physically stop an individual from entering a restricted area or prevent tailgating in real-time.
• B. A mantrap (also known as a security vestibule) is specifically designed to prevent tailgating by using interlocking doors that only allow one person to pass through at a time. Integrating biometric authentication ensures that the individual inside the mantrap is verified using unique physical characteristics, making it the most effective preventive control for high-security areas like AI server rooms.
• C. RFID badge readers provide a basic level of access control but are susceptible to theft, cloning, or sharing. Furthermore, placing them at the main building entrance does not provide the granular security required for a specific internal server room, and they do not physically prevent multiple people from walking through a door once it is opened.
• D. Motion sensors are detective controls that can trigger alarms or record events when movement is detected in a restricted area. They are effective for monitoring after-hours activity but provide no physical barrier to prevent unauthorized entry or tailgating.

Correct answer: A — Model watermarking

• A. Correct. Model watermarking is a technique used to embed a unique, often imperceptible, and verifiable signature into an AI model's parameters or behavior. This allows the organization to demonstrate provenance and prove ownership if the model is stolen, copied, or deployed in an unauthorized environment.
• B. Incorrect. Homomorphic encryption is a cryptographic method that allows for computation on encrypted data without needing to decrypt it first. While it protects the confidentiality of data during processing, it does not provide a mechanism for establishing ownership of the model itself.
• C. Incorrect. Differential privacy is a technique used to protect the privacy of individual data points within a training dataset by adding mathematical noise. It helps prevent membership inference attacks and data leakage but does not help in proving intellectual property ownership.
• D. Incorrect. Federated learning is a decentralized machine learning approach where models are trained across multiple edge devices or servers holding local data samples without exchanging the raw data. This enhances data privacy and locality but does not embed ownership evidence into the resulting model.

Correct answer: B — Impossible travel

• A. Incorrect. Privilege escalation involves an attacker or user gaining higher permissions or access rights (e.g., moving from a standard user to a local admin) than they are authorized to have. This scenario focuses on authentication location rather than permission changes.
• B. Correct. Impossible travel is a specific UEBA detection where a single account authenticates from two geographically distant locations within a timeframe that is physically impossible for a person to travel. Logging in from New York and Tokyo within 15 minutes is a classic indicator of credential compromise or VPN abuse.
• C. Incorrect. Lateral movement occurs when an attacker moves through a network from system to system to locate high-value targets. While it may follow an initial login, the specific anomaly triggered by the geographic distance and timing is classified as impossible travel.
• D. Incorrect. Data exfiltration refers to the unauthorized transfer of data out of an organization's environment. The scenario describes a suspicious login pattern, not the actual transfer or theft of data.

Correct answer: A — False positive reduction

• A. Correct. Reducing alerts for benign vulnerability scanner traffic is a classic example of false positive reduction. By using AI to distinguish harmless, routine activity from truly suspicious behavior, the SIEM lowers alert fatigue and allows security analysts to focus their limited resources on genuine threats.
• B. Incorrect. False negative elimination would mean the system is identifying malicious activities that were previously missed. The scenario specifically describes the reduction of noise from benign sources, which is a reduction in false alarms, not the discovery of missed threats.
• C. Incorrect. Increased log retention refers to the capacity and duration for which historical log data is stored. While AI might help analyze these logs, the specific operational benefit described here is about alert quality, not storage capacity.
• D. Incorrect. Automated patching involves the programmatic application of software updates to fix vulnerabilities. The benefit described in the scenario is related to alert filtering and prioritization within the SIEM, not the remediation of vulnerabilities.

Correct answer: A — alert fatigue

• A. Correct. Alert fatigue describes the exhaustion and desensitization experienced by security analysts when faced with a high volume of low-fidelity alerts or false positives. By using AI to filter noise and prioritize high-confidence threats, the SOC improves the signal-to-noise ratio, allowing analysts to focus on meaningful incidents and reducing the risk of burnout.
• B. Incorrect. Network latency refers to the delay in data transmission across a network. While AI can be used for network optimization, the suppression and escalation of security alerts in a SOC workflow do not address communication delays.
• C. Incorrect. Encryption overhead refers to the computational resources (CPU and time) required to perform cryptographic operations. Alert triage and management are operational security tasks and are unrelated to the performance costs associated with data encryption.

Correct answer: B — Eradication

• A. Preparation involves establishing infrastructure, procedures, and training before an incident occurs. While the SOAR platform itself is set up during preparation, the active removal of a threat is a response phase action.
• B. Eradication is the phase where the malicious threat or artifact is permanently removed from the environment. Deleting a confirmed phishing email from user inboxes and checking malicious URLs against threat feeds are actions specifically designed to eliminate the threat's presence and its root cause.
• C. Recovery focuses on restoring systems to normal operation and ensuring business continuity after a threat has been handled. Deleting a malicious artifact is considered part of the cleanup (eradication) rather than the restoration of services.
• D. Post-incident activity involves reviewing the incident, documenting lessons learned, and improving future response capabilities. The actions described in the scenario are immediate response steps taken during the incident, not the retrospective analysis performed after resolution.

Correct answer: B — Predictive analytics

• A. Retrospective forensics involves analyzing past security incidents after they have occurred to determine root causes and impact. It focuses on historical investigation rather than forecasting future events.
• B. Predictive analytics uses machine learning, statistical algorithms, and historical/current data to identify the likelihood of future outcomes. By correlating network vulnerabilities, threat intelligence, and industry context to forecast future targets, the platform is performing predictive modeling.
• C. Automated containment refers to automated actions taken to isolate or mitigate a threat once it has been detected, such as quarantining an infected endpoint. This is a reactive response capability, not a forecasting capability.
• D. Data loss prevention (DLP) consists of strategies and tools designed to detect and prevent unauthorized data exfiltration. While it protects sensitive assets, it does not typically involve forecasting which assets are likely to be targeted based on external threat trends.

Correct answer: A — Vishing using voice cloning

• A. Correct. Vishing (voice phishing) using voice cloning involves generating synthetic speech to impersonate a trusted individual, such as a CEO, to deceive a target. The detection of unnatural phonetic artifacts and synthetic frequency patterns by the AI tool are definitive indicators of AI-generated or cloned audio.
• B. Incorrect. Spear phishing via SMS (often called smishing) involves targeted text messages sent to specific individuals. This scenario involves an audio voicemail, which is the hallmark of vishing rather than SMS-based phishing.
• C. Incorrect. A Man-in-the-Middle (MitM) attack involves an attacker intercepting or altering the communication between two parties on a network. This scenario involves the analysis of synthetic media artifacts in a voicemail, not the interception of a live communication stream.
• D. Incorrect. Cross-Site Request Forgery (CSRF) is a web application vulnerability that forces an authenticated user to execute unwanted actions on a web application. It has no relation to audio synthesis, voicemail, or social engineering via phone.

Correct answer: B — Contextualizing logins with historical user behavior to filter out benign travel and VPN usage.

• A. Incorrect. Automatically blocking all IP addresses from foreign countries is a crude, rule-based approach rather than an intelligent triage solution. This action would likely cause significant business disruption and create many false positives for legitimate users traveling abroad.
• B. Correct. AI-assisted triage tools utilize behavioral baselining to compare current login events against historical patterns, typical locations, and known VPN usage. By contextualizing these alerts, the AI can distinguish between legitimate user activity and actual threats, significantly reducing alert fatigue and prioritizing true anomalies for the analyst.
• C. Incorrect. While encrypting authentication logs is a standard security practice for protecting data integrity and confidentiality, it does not assist in the analysis, triage, or contextualization of alerts to reduce the workload of a SOC analyst.
• D. Incorrect. Replacing an identity provider with a blockchain solution is a fundamental infrastructure change that does not address the core problem of triaging existing SIEM alerts. It is unrelated to the AI's ability to improve decision-making during the incident response process.

Correct answer: B — Natural Language Processing (NLP)

• A. Generative Adversarial Networks (GANs) are primarily used to generate synthetic data, such as images or text, through the competition of two neural networks. They are not designed for extracting semantic meaning from unstructured text or correlating specific alerts with external documentation.
• B. Natural Language Processing (NLP) is specifically designed to analyze and interpret human language. It is the ideal capability for processing unstructured text from sources like PDFs and blogs, allowing the SOC to extract relevant entities (such as IoCs or TTPs) and map them to SIEM alerts.
• C. Reinforcement Learning focuses on training agents to make a sequence of decisions based on a system of rewards and punishments. It is used for optimization and control tasks, not for the linguistic analysis required to understand threat intelligence reports.
• D. K-Means Clustering is an unsupervised machine learning algorithm used to group data points based on numerical feature similarity. It lacks the inherent linguistic understanding necessary to parse unstructured security reports or perform context-aware correlation between alerts and text.

Correct answer: B — Implementing a continuous feedback loop to retrain the model on recent threat data to combat model drift.

• A. Incorrect. Reverting SIEM software to a previous backup addresses software stability or configuration issues, but it does not fix an AI model's failure to recognize new threat patterns. The issue is a mismatch between the model's training data and current real-world data, not a software version error.
• B. Correct. This scenario describes model drift (specifically concept drift), where the statistical properties of the target variable change over time as adversaries evolve. Implementing a continuous feedback loop allows the model to be regularly retrained with up-to-date threat intelligence and recent samples, ensuring it maintains accuracy against new variants.
• C. Incorrect. Increasing hardware resources like CPU and RAM may improve the speed (latency) or throughput of the model, but it does not improve the quality of its predictions. If the model logic is stale, more computational power will only produce incorrect results faster.
• D. Incorrect. Switching to a linear regression model would likely decrease performance. Neural networks are far better at identifying complex, non-linear patterns found in modern ransomware. The problem is not the algorithm type, but the fact that the algorithm was trained on outdated data.

Correct answer: A — The ability to opt-out of automated decision-making and profiling.

• A. Correct. The CPRA (which expanded the CCPA) specifically grants consumers the right to opt-out of the use of automated decision-making technology (ADMT) and profiling. This includes profiling for targeted behavioral advertising, which is a core provision designed to give consumers control over how their personal information is processed by AI systems.
• B. Incorrect. The CPRA focuses on transparency, access, and control rights rather than financial remuneration. There is no requirement for companies to provide monetary compensation to consumers for using their data in AI training models.
• C. Incorrect. While the CPRA requires transparency regarding the logic involved in automated decision-making, it does not mandate that companies release their proprietary algorithms or source code as open-source.
• D. Incorrect. The CPRA generally follows an opt-out model for most data processing activities. While mandatory opt-in consent is required in specific cases (such as for the data of minors or certain uses of sensitive personal information), it is not the universal standard for all retail data collection and profiling.

Correct answer: C — ISO/IEC 42001

• A. ISO/IEC 27001 is the international standard for Information Security Management Systems (ISMS). It focuses on protecting the confidentiality, integrity, and availability of information assets in general, but it is not specifically tailored to the unique lifecycle and ethical risks of AI.
• B. ISO/IEC 27701 is an extension to ISO/IEC 27001 focused on Privacy Information Management Systems (PIMS). While it is critical for handling personal data, it does not provide the comprehensive governance framework required for an AI Management System (AIMS).
• C. ISO/IEC 42001 is the specific international standard designed for Artificial Intelligence Management Systems (AIMS). It provides a certifiable framework for organizations to govern AI responsibly, manage AI-specific risks, and ensure transparency and accountability throughout the AI lifecycle.
• D. ISO 9001 is the general standard for Quality Management Systems (QMS). While it ensures that organizations meet customer and regulatory requirements through quality processes, it is not specialized for AI governance or AI-related risk management.

Correct answers: B, D — It empowers existing sector-specific regulators (e.g., FCA, ICO) to apply cross-sectoral AI principles.; It focuses on a pro-innovation framework rather than immediate heavy statutory regulation.

• A. The UK has deliberately avoided creating a single, centralized AI regulator. Instead, it utilizes a decentralized, sector-based approach that relies on the expertise of existing bodies to oversee AI within their specific jurisdictions.
• B. This accurately describes the UK's 'sector-led' approach. It empowers existing regulators, such as the Financial Conduct Authority (FCA) and the Information Commissioner's Office (ICO), to apply cross-sectoral principles—like safety, transparency, and fairness—within their own domains.
• C. The UK framework is distinct from the European Union's approach. While the EU has implemented a strict, horizontal AI Act with risk-based tiers, the UK has opted for a more flexible, non-statutory, and principles-based framework.
• D. The UK's strategy, as outlined in its official white papers, prioritizes a pro-innovation environment. It aims to support AI development and economic growth by using existing laws and regulatory powers to manage risks, rather than introducing immediate, heavy-handed statutory regulation.
• E. There is no blanket ban on generative AI in these sectors. Instead, the UK encourages responsible innovation, with regulators providing sector-specific guidance on risk management and governance for AI applications.

Correct answer: B — Negotiate a contract clause explicitly prohibiting the use of API inputs for vendor model training

• A. Encrypting data at rest on the local network protects the data while stored internally, but it does not address how the third-party vendor processes or utilizes the data once it has been transmitted to their service via the API.
• B. Negotiating a contract clause or reviewing the Data Processing Agreement (DPA) is the most effective GRC-driven mitigation. This provides a legal and governance framework that explicitly prohibits the vendor from using the organization's proprietary data for model training, which is a standard requirement for enterprise-grade AI integrations.
• C. Rate limiting is a technical control used to manage traffic volume, prevent denial-of-service (DoS) attacks, or control costs. It has no bearing on the vendor's data usage policies or their ability to ingest received data into training sets.
• D. While anonymization can reduce the sensitivity of the data, it is often technically difficult to fully de-identify complex financial reports without losing context. Furthermore, this is a secondary technical control that does not replace the primary governance requirement of establishing legal data-use restrictions with the vendor.

Correct answer: A — AI Governance Board

• A. An AI Governance Board provides strategic oversight, evaluates enterprise-wide risks, and ensures alignment with corporate values and ethics across all AI initiatives. It is a cross-functional body essential for standardizing AI adoption, policy development, and managing the multi-faceted risks (legal, ethical, technical) associated with AI tools, whether they are procured (SaaS) or built in-house.
• B. A Security Operations Center (SOC) is an operational unit focused on real-time monitoring, detection, and response to security threats. While it might monitor AI-based security tools or respond to AI-related security events, it lacks the mandate for high-level governance, strategic risk evaluation, or aligning AI usage with corporate values.
• C. An Incident Response Team is reactive, tasked with mitigating and recovering from security or operational incidents after they occur. It does not provide the ongoing strategic oversight, risk assessment, or policy frameworks required to govern enterprise-wide AI adoption proactively.
• D. A Change Advisory Board (CAB) reviews and manages the risk associated with changes to IT services and infrastructure. While a CAB might review the technical deployment of an AI application, it is not structured to handle the broader strategic, ethical, and organizational alignment risks inherent in an enterprise AI strategy.

Correct answers: A, B, D — Map; Measure; Govern

• A. Correct. 'Map' is one of the four core functions of the NIST AI RMF. It focuses on establishing the context, identifying the intended use, and characterizing potential impacts and risks of an AI system.
• B. Correct. 'Measure' is a core function in the NIST AI RMF. It involves using quantitative, qualitative, or mixed-method tools to analyze, assess, and track AI risks and related performance characteristics.
• C. Incorrect. While risk mitigation is a key goal of risk management, 'Mitigate' is not a standalone core function in the NIST AI RMF. Treatment and mitigation actions are performed under the 'Manage' core function.
• D. Correct. 'Govern' is the foundation of the NIST AI RMF. It is a cross-cutting function that establishes organizational culture, policies, and accountability to manage AI risks effectively.
• E. Incorrect. 'Recover' is a core function of the NIST Cybersecurity Framework (CSF), but it is not one of the four core functions defined in the NIST AI Risk Management Framework (which are Govern, Map, Measure, and Manage).

Correct answer: B — Fairness

• A. Transparency refers to the openness and clarity regarding how an AI system operates, its data sources, and its decision-making processes. While important for building trust, the primary focus of this scenario is on equitable model performance rather than explainability or openness.
• B. Fairness in AI involves ensuring that systems provide equitable outcomes and do not discriminate against individuals or groups based on characteristics such as race or skin tone. By identifying accuracy disparities and retraining the model with a more diverse dataset to eliminate bias, the team is directly addressing the principle of fairness.
• C. Privacy pertains to protecting personal information and ensuring data confidentiality. While privacy is critical in healthcare AI, the issue here involves model bias and performance inequality rather than the exposure or misuse of sensitive patient data.
• D. Accountability refers to assigning responsibility for the AI system's impacts and ensuring there is oversight for its decisions. While the team is demonstrating responsible oversight by halting deployment, the specific ethical gap they are resolving is the biased performance, which falls under the category of fairness.

Correct answer: B — Intellectual property infringement

• A. Incorrect. Data sovereignty refers to the legal jurisdiction and requirements over data based on its geographic location (where it is stored or processed). This scenario focuses on the ownership and rights of the content itself, not jurisdictional data laws.
• B. Correct. Intellectual property (IP) infringement is the primary risk here. Training an AI model on copyrighted images without obtaining consent or a valid license from the original creators can lead to copyright violations, legal liability, and ethical concerns regarding the fair use of creative works.
• C. Incorrect. Model inversion attacks are a security threat where an adversary attempts to reconstruct training data or sensitive information by querying the model. While it involves training data, the legal team's concern is about the unauthorized sourcing of data, not its extraction by attackers.
• D. Incorrect. Data poisoning is an adversarial attack where malicious data is injected into a training set to manipulate the model's behavior or accuracy. The legal issue in this scenario is about improper sourcing and lack of permission for legitimate data, not malicious tampering.