Hard Difficulty Questions

Vertex AI Pipelines caching

Why wrong: Caching optimizes performance, not security auditability.

Cloud Data Loss Prevention (DLP) API

Why wrong: DLP is for data classification and redaction, not lineage/provenance.

Cloud Armor WAF

Why wrong: Cloud Armor protects endpoints, not training data integrity.

Allowing only one developer to modify the model.

Why wrong: This creates a single point of failure and is not a version control practice.

Storing only the production model artifact.

Why wrong: You must store all versions to allow for rollbacks and investigation.

Auditing is only for small companies.

Why wrong: Auditing is required for all scales.

Monitoring is only for external systems.

Why wrong: Monitoring applies to all systems.

Continuous monitoring is cheaper.

Why wrong: Continuous is usually more resource-intensive.

Using a low-latency network.

Why wrong: Irrelevant to risk.

Using a well-documented framework.

Why wrong: Decreases risk.

Using a louder cooling system.

Why wrong: Irrelevant.

Increasing the model's training duration.

Why wrong: Does not prevent data leakage.

To gain access to the raw training database.

Why wrong: This is a data exfiltration attack.

To cause the model to crash.

Why wrong: This is a DoS attack.

To inject malicious weights into the model.

Why wrong: This is a poisoning attack.

It uses a secondary model to detect malicious input.

Why wrong: That is an input detector, not adversarial training.

It restricts input to a whitelist of authorized users.

Why wrong: That is access control.

It adds noise to the weights of the model.

Why wrong: That is weight perturbation, not adversarial training.

The model design and architecture selection.

Why wrong: This is the research phase, not the ingestion phase.

Model monitoring and logging.

Why wrong: This happens after the model is trained.

Inference and deployment.

Why wrong: This is the stage for evasion attacks, not poisoning.

The office lease agreement.

Why wrong: Irrelevant.

A copy of the employee handbook.

Why wrong: Irrelevant.

The model's favorite color.

Why wrong: Irrelevant.

The developer's favorite food.

Why wrong: Irrelevant.

Requirements to publish all internal vulnerabilities publicly.

Why wrong: Public disclosure is not required in an internal IR plan.

Automated deletion of all logs.

Why wrong: Logs are essential for forensic investigation.

Overfitting of the LLM on the training set.

Why wrong: This happens during training, not during retrieval.

Gradient-based evasion attacks on the LLM.

Why wrong: This is a standard adversarial ML attack, not RAG-specific.

Prompt injection targeting the LLM model weights.

Why wrong: This is a standard LLM vulnerability, not specific to the RAG retrieval mechanism.

Access to all employees' private emails.

Why wrong: This is a privacy and legal violation.

A mandate to use the most expensive GPUs available.

Why wrong: This is a cost issue, not governance.

It is the slow down of the model due to heavy load

Why wrong: This is a performance/DoS issue.

It is the unauthorized access to the server room

Why wrong: This is a physical security incident.

It is the accidental deletion of model files

Why wrong: This is a data loss incident.

Brute forcing the user password

Why wrong: This is an IT authentication attack.

Adding noise to the output

Why wrong: This is a defense against privacy attacks.

Stealing the model's entire weights file

Why wrong: This is model extraction, not membership inference.

The names of all employees who use the AI.

Why wrong: User lists are not part of an MRM policy.

A list of all model research papers published by the company.

Why wrong: Research history is not a risk management policy requirement.

Asking the vendor to delete all their internal logs.

Why wrong: This destroys evidence and accountability.

Requiring all employees to use the vendor's internal Wi-Fi.

Why wrong: This is not a valid risk control.

Requiring all models to be open-source.

Why wrong: Open-source is not a requirement for risk management.

Implementing a strict 'no external API' policy for all models.

Why wrong: This is overly restrictive and not a standard NIST requirement.

It is required to make the model run faster.

Why wrong: XAI often adds overhead, not speed.

It encrypts the input data.

Why wrong: XAI has nothing to do with encryption.

It enables models to train without labels.

Why wrong: XAI is not a training method.

To enable cross-cloud model training

Why wrong: It is not related to cloud infrastructure topology.

To allow for better interpretation of model layers

Why wrong: It is a privacy tool, not an interpretability tool.

To ensure the model is faster at inference time

Why wrong: It often adds computational overhead.