Hard Difficulty Questions

The color of the server racks

Why wrong: Not a model risk.

The speed of the local internet connection

Why wrong: Not a model risk.

The programming language used by the vendor

Why wrong: The language is irrelevant to governance risk.

Vendor's model update and retraining frequency

Why wrong: This is an operational detail, not an inherent governance concern compared to data handling.

The physical location of the vendor's corporate office

Why wrong: While relevant to business continuity, it is not a direct AI governance risk.

Neural network weight pruning

Why wrong: Specific to neural networks.

Linear regression coefficients

Why wrong: Specific to linear models.

Checking if the server is plugged in

Why wrong: Basic hardware checks are not adversarial testing.

Checking the model's training accuracy

Why wrong: Training accuracy does not measure adversarial robustness.

Asking the team to describe their favorite movies

Why wrong: Casual conversation is not a security test.

Color of the UI

Why wrong: Irrelevant.

Number of coffee machines

Why wrong: Irrelevant.

Ignore the change

Why wrong: Negligent audit practice.

Delete the model

Why wrong: Premature.

Upgrade the hardware

Why wrong: Unrelated.

Increasing compute power

Why wrong: Irrelevant to privacy.

Model quantization

Why wrong: Optimization task.

Removing the model

Why wrong: Too extreme.

Whether the team uses a specific cloud provider

Why wrong: The choice of cloud vendor is not a governance requirement.

The speed of the training job

Why wrong: Performance speed is operational, not a governance/compliance requirement.

Usage of proprietary algorithms only

Why wrong: Proprietary status does not equate to governance compliance.

To encrypt model weights in memory

Why wrong: Shadow mode has nothing to do with encryption.

To hide model artifacts from unauthorized users

Why wrong: Shadow mode is a testing technique, not a security control.

To automate hyperparameter optimization

Why wrong: This is a model development activity.

The office location of the data scientists

Why wrong: Not relevant to technical replication.

The brand of the coffee machine used

Why wrong: Irrelevant.

The developer's personal email address

Why wrong: Sensitive and unnecessary for replication.

Reducing inference latency

Why wrong: This is a performance goal.

Increasing training bias

Why wrong: This is an unintended outcome, not an attack.

Overfitting the data

Why wrong: Overfitting is a model state, not an attack type.

Running unit tests for code

Why wrong: Unit tests don't check for semantic jailbreaks.

Reviewing server logs

Why wrong: Logs show what happened, but not vulnerabilities before they occur.

Checking network bandwidth

Why wrong: Bandwidth is unrelated to safety filters.

The font size used in the report

Why wrong: Irrelevant.

The names of all employees who disagreed

Why wrong: Unprofessional and irrelevant.

Increased training latency

Why wrong: This is a performance risk, not a fundamental security or governance risk.

Lack of access to training hardware

Why wrong: Hardware access is a capacity planning issue, not an architectural risk.

Centralized data breaches

Why wrong: Federated learning is designed to avoid centralizing data, reducing this specific risk.

Model training speed

Why wrong: Epsilon is for privacy, not performance.

Learning rate scheduling

Why wrong: Learning rate is a hyperparameter for convergence.

Database query limit

Why wrong: This relates to API rate limiting.

The physical location of the cloud server hosting the model

Why wrong: Server location is an infrastructure property, not a model performance indicator.

The number of lines of code in the model deployment script

Why wrong: Code volume is not a proxy for model reliability.

The number of users accessing the model API

Why wrong: User count is a usage metric, not a reliability or performance metric.

Lack of a social media presence for the company

Why wrong: Social media presence is irrelevant to security.

Using a low-cost cloud provider

Why wrong: Provider cost does not cause security vulnerabilities.

Failure to update the company logo

Why wrong: Branding has no effect on model security.

Data distribution drift

Why wrong: Drift is measured via statistical tests.

Overall model performance

Why wrong: LIME is for local, not global analysis.

Adversarial robustness

Why wrong: LIME is not an adversarial testing tool.

The number of lines of code in the pipeline

Why wrong: Code length is not a security metric.

Office access badge logs

Why wrong: Physical security is a general IT control, not pipeline-specific.

Increasing the number of neurons in the hidden layers

Why wrong: Adding capacity doesn't solve drift; it just makes the model more complex.

Moving the model to a larger GPU cluster

Why wrong: Infrastructure upgrades don't fix the underlying statistical issue of concept drift.

Restricting access to the training dataset

Why wrong: Access restriction is a security control, not a drift mitigation tactic.