AI0-001 AI Security, Ethics and Governance Practice Test 3 — 15 Questions

Question 1

A security analyst notices that an AI model used for facial recognition is returning unusually high confidence scores for certain individuals while consistently misidentifying others. Which type of attack is most likely occurring?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Model inversion attack. A model inversion attack allows an adversary to reconstruct training data or infer sensitive attributes from the model's outputs. In this scenario, the unusually high confidence scores for certain individuals and misidentification of others indicate that the attacker is exploiting the model's internal representations to extract information about the training data, leading to biased or overconfident predictions for specific classes.

Answer

Data poisoning

Answer

Evasion attack

Answer

Model extraction attack

Question 2

A bank deploys an AI system to approve loan applications. During testing, the model denies a disproportionate number of applicants from a particular demographic group, even after controlling for credit history. Which ethical principle is being violated?

Accepted Answer

Fairness. The AI system's disparate impact on a demographic group, even after controlling for credit history, directly violates the principle of fairness. Fairness in AI requires that models do not produce biased outcomes that systematically disadvantage protected groups, regardless of whether the bias stems from training data, feature selection, or algorithmic design. This scenario describes a clear case of algorithmic bias, which fairness principles aim to prevent.

Answer

Transparency

Answer

Privacy

Answer

Accountability

Question 3

A healthcare organization uses an AI model to predict patient readmission risk. To comply with patient privacy regulations, they apply differential privacy during training. What is the primary trade-off of using differential privacy?

Accepted Answer

Reduced model accuracy for increased privacy. Differential privacy works by adding calibrated noise to the training process or model outputs, which directly reduces the model's accuracy in exchange for a quantifiable privacy guarantee (e.g., ε-differential privacy). This trade-off is fundamental: stronger privacy (lower ε) requires more noise, which degrades predictive performance. The healthcare organization must balance the need to protect patient data against the clinical utility of accurate readmission predictions.

Answer

Increased training time for reduced bias

Answer

Lower interpretability for higher fairness

Answer

Faster inference for lower memory usage

Question 4

A cybersecurity analyst monitors an AI chatbot that frequently produces offensive responses when given specific prompts. The development team suspects an adversarial attack. Which mitigation strategy is most effective against such prompt injection attacks?

Accepted Answer

Implement input validation and sanitization. Prompt injection attacks exploit the model's inability to distinguish between user input and system instructions. Input validation and sanitization (e.g., filtering special characters, enforcing strict schema checks, and using allowlists) directly neutralize malicious payloads before they reach the model's inference engine, preventing the model from executing unintended commands.

Answer

Retrain the model on a larger, curated dataset

Answer

Encrypt all communication between users and the chatbot

Answer

Reduce the model's number of parameters

Question 5

A company's AI governance board requires each model to have a model card documenting intended use, performance metrics, and limitations. What is the primary purpose of a model card?

Accepted Answer

To provide transparent documentation of model capabilities and limitations. A model card is a standardized documentation framework that provides transparent, concise information about a machine learning model's intended use, performance metrics, and limitations. This transparency enables stakeholders to understand the model's capabilities and potential biases, ensuring responsible deployment and governance as required by AI ethics and governance frameworks.

Answer

To specify the exact training algorithm and hyperparameters

Answer

To outline a complete risk assessment framework

Answer

To serve as a legal contract between developers and users

Question 6

An AI system in a self-driving car misinterprets a stop sign due to a small sticker placed on it. This is an example of which security vulnerability?

Accepted Answer

Adversarial example attack. The sticker on the stop sign creates a small perturbation that causes the AI model's image classifier to misclassify the sign (e.g., as a speed limit sign). This is the defining characteristic of an adversarial example attack, where crafted input perturbations exploit model vulnerabilities to cause incorrect predictions.

Answer

Supply chain attack

Answer

Model inversion attack

Answer

Data poisoning attack

Question 7

Which TWO of the following are effective techniques for detecting bias in an AI model?

Accepted Answer

Fairness metrics such as equal opportunity difference. Fairness metrics such as equal opportunity difference directly quantify bias by measuring the difference in true positive rates between privileged and unprivileged groups. A value of zero indicates perfect fairness, while non-zero values reveal disparate treatment, making it a standard technique for bias detection in AI models.

Answer

Feature importance scores

Answer

Confusion matrix on the entire dataset

Answer

Cross-validation accuracy

Question 8

Which THREE of the following are key components of an AI governance framework?

Accepted Answer

Risk assessment and mitigation plans. Risk assessment and mitigation plans are a core component of an AI governance framework, ensuring that potential harms, biases, and security vulnerabilities are identified and addressed before deployment. This aligns with frameworks like NIST AI RMF, which mandates continuous risk monitoring and mitigation strategies to maintain ethical and secure AI operations.

Answer

Cloud infrastructure configuration

Answer

Model accuracy benchmarks

Question 9

Which TWO of the following are common threats to AI model security?

Accepted Answer

Data poisoning. B is correct because data poisoning involves an attacker injecting malicious data into the training set to corrupt the model's learning process, causing it to make incorrect predictions or classifications. This directly undermines the integrity of the AI model by manipulating its foundational training data.

Answer

SQL injection

Answer

Distributed denial-of-service (DDoS)

Answer

Phishing attacks

Question 10

Refer to the exhibit. A security analyst reviews the monitoring log for an AI fraud detection model. Which of the following is the most likely cause of the multiple alerts?

Accepted Answer

An adversarial attack attempt. An adversarial attack attempt is the most likely cause because the monitoring log shows multiple alerts triggered by subtle, crafted perturbations in input data designed to cause the AI fraud detection model to misclassify legitimate transactions as fraudulent or vice versa. Unlike data poisoning, which corrupts the training dataset over time, adversarial attacks target the model's inference phase, exploiting its sensitivity to small input variations to produce incorrect outputs without altering the underlying training data.

Answer

Data poisoning of the training dataset

Answer

A network hardware failure

Answer

A scheduled model retraining process

Question 11

Refer to the exhibit. A system administrator sees these logs from an AI inference pipeline. What is the most likely sequence of events?

Accepted Answer

An adversarial input was blocked by the security filter. The logs show the security filter actively blocking an input (e.g., 'Blocked by security filter: input contains adversarial pattern') and then the pipeline continuing normally. This indicates the filter detected and stopped an adversarial input before it reached the model, preventing any error or corruption. The absence of NaN or model errors confirms the filter functioned as designed.

Answer

Data poisoning corrupted the model, causing NaN outputs

Answer

The security filter failed to detect an attack and the model returned an error

Answer

A non-adversarial input caused a NaN error due to missing data

Question 12

A security team discovers that an AI-based anomaly detection system frequently misclassifies benign network traffic as malicious when the source IP is from a specific geographic region. Which type of AI vulnerability is most likely being exploited?

Accepted Answer

Adversarial evasion. The scenario describes an AI-based anomaly detection system that misclassifies benign traffic from a specific geographic region as malicious. This is a classic example of an adversarial evasion attack, where an attacker crafts inputs (in this case, network traffic) that appear benign to human analysts but cause the AI model to misclassify them. The geographic bias suggests the attacker is exploiting the model's learned decision boundary, likely by manipulating features such as source IP or packet timing to evade detection.

Answer

Data poisoning

Answer

Model inversion

Answer

Membership inference

Question 13

A healthcare organization uses an AI model to recommend treatment plans. The model was trained on data from a single hospital, and now treats patients from multiple demographics. Which ethical concern is most critical?

Accepted Answer

Fairness and bias in predictions. The model was trained on data from a single hospital, which likely has a homogeneous demographic profile. When deployed across multiple demographics, the model may produce biased or unfair predictions for underrepresented groups, making fairness and bias the most critical ethical concern. This directly violates the principle of distributive justice in AI ethics.

Answer

Accountability for treatment outcomes

Answer

Lack of transparency in model decisions

Answer

Privacy violations in training data

Question 14

An AI system used for resume screening is found to consistently reject female candidates for technical roles. The data science team retrains the model after removing the 'gender' feature, but the bias persists. What is the most likely cause?

Accepted Answer

The model uses proxy variables that correlate with gender. Even after removing the explicit 'gender' feature, the model can still learn biased patterns from proxy variables that correlate strongly with gender, such as years of experience (which may be lower for women due to career breaks), educational institutions attended, or even hobbies listed on resumes. These proxies act as surrogates for the protected attribute, allowing the model to effectively discriminate despite the feature being removed. This is a well-known phenomenon in algorithmic fairness called 'redundant encoding' or 'proxy discrimination.'

Answer

The model architecture is too complex

Answer

The training data still contains historical hiring bias

Answer

The evaluation metric does not measure fairness

Question 15

A company implements an AI-based chatbot for customer service. After deployment, customers report that the chatbot sometimes uses offensive language. The development team reviews the training data and finds no explicit offensive content. What is the most likely explanation?

Accepted Answer

The model learned biased language patterns from the training corpus. The chatbot's offensive language likely stems from biased or toxic patterns present in the training corpus, even if no explicit offensive content was flagged. Large language models learn statistical associations from their training data, and if the corpus contains subtle biases, stereotypes, or indirect toxic language, the model can reproduce these patterns in its responses. This is a well-known issue in AI ethics and governance, where models inadvertently amplify societal biases embedded in the data.

Answer

There is a bug in the deployment pipeline

Answer

The model is overfitting to rare examples

Answer

The training data was poisoned by an attacker