AAISM AI Technologies And Controls — 20 Questions

Question 1

Which TWO actions should be taken to enforce Least Privilege access for developers interacting with SageMaker Model Monitor?

Accepted Answer

Restrict access to the 'sagemaker:CreateModelMonitor' action. Access to Model Monitor requires permissions to both the infrastructure (S3) and the specific model monitoring job definitions.

Answer

Allow all 'iam:PassRole' actions

Answer

Assign the 'SageMakerFullAccess' managed policy

Answer

Enable root access to the underlying EC2 instance

Question 2

You are configuring AWS Bedrock Guardrails to prevent the model from returning PII. Which specific configuration setting should be enabled to ensure PII filtering is applied to both prompt and response?

Accepted Answer

Enable PII filtering in the Policy Settings for both 'input' and 'output' content flows.. AWS Bedrock Guardrails allow for PII entity redaction or blocking; the configuration must be set to filter both incoming prompts and outgoing model responses to ensure comprehensive protection.

Answer

Use the prompt engineering instruction 'Do not output PII' in the System Prompt field.

Answer

Set the content moderation filter level to 'High' in the model configuration.

Answer

Apply an IAM policy to the Bedrock execution role restricting access to PII-containing datasets.

Question 3

You are configuring Microsoft Defender for Cloud to monitor your Azure OpenAI service. Which specific resource diagnostic setting must be enabled to ensure that prompt injection attempts are logged in the Log Analytics workspace?

Accepted Answer

Enable 'RequestResponse' diagnostic logs for the AI resource. To detect security threats like prompt injection, you must enable diagnostic logging for the Azure OpenAI service specifically for the 'RequestResponse' category, which captures the interaction payload.

Answer

Enable 'AccessControl' logging in Entra ID

Answer

Enable 'MetricAlerts' at the subscription level

Answer

Enable 'AuditLogs' in the Storage Account settings

Question 4

During an adversarial training cycle, you notice the model's accuracy on the validation set drops significantly despite high training accuracy. Which technique should you apply to mitigate potential model poisoning?

Accepted Answer

Implement Differential Privacy (DP-SGD). Differential privacy helps ensure that the model does not memorize specific training data samples, which is a common target for poisoning attacks aimed at steering model output.

Answer

Increase the learning rate

Answer

Enable Model Versioning in MLflow

Answer

Switch to a pre-trained Transformer architecture

Question 5

When implementing Google Cloud's 'Vertex AI' model guardrails, you need to prevent the leakage of PII in model outputs. Which configuration tool is used to enforce these constraints at the request/response level?

Accepted Answer

Vertex AI 'Safety Settings' filters. The Vertex AI 'Safety Settings' and 'Content filtering' allow practitioners to define thresholds for harassment, hate speech, and PII leakage.

Answer

Vertex AI 'Feature Store' access control

Answer

Vertex AI 'Model Registry' IAM roles

Answer

Vertex AI 'Pipeline' execution logs

Question 6

You are configuring Microsoft Defender for Cloud for an Azure OpenAI deployment. To detect prompt injection attempts effectively, which specific setting must be enabled in the 'Microsoft Defender for Cloud Apps' dashboard?

Accepted Answer

Enable 'AI Security Posture Management' (AI-SPM). Microsoft Defender for Cloud leverages 'AI Security Posture Management' to identify anomalous prompts that resemble injection patterns.

Answer

Set 'Log Analytics' retention to 365 days

Answer

Configure 'Conditional Access' policies for API keys

Answer

Enable 'Azure Policy' for resource tagging

Question 7

An organization is using Google Cloud Vertex AI and suspects an adversarial poisoning attack on a custom training dataset. Which tool should be used to verify the integrity and provenance of the training artifacts?

Accepted Answer

Vertex AI ML Metadata. Vertex AI Model Garden and Metadata are used for lineage tracking, but specifically for provenance and audit, the Vertex AI ML Metadata service allows tracking of dataset versions and training pipeline execution history.

Answer

Vertex AI Pipelines caching

Answer

Cloud Data Loss Prevention (DLP) API

Answer

Cloud Armor WAF

Question 8

You are auditing access controls for a Microsoft Azure OpenAI deployment. Which TWO of the following configurations are required to ensure identity-based access and logging for model inference?

Accepted Answer

Enable 'Diagnostic settings' to route logs to a Log Analytics workspace.. Azure OpenAI requires Entra ID (Azure AD) for authentication and Azure Monitor/Log Analytics for auditing request/response patterns.

Answer

Enable the 'Prompt Injection' toggle in the Azure Portal global settings.

Answer

Create a shared API key for all users to simplify service account management.

Answer

Set the model to 'Public Access' in the Networking tab for easier debugging.

Question 9

You are managing a PyTorch model and suspect data poisoning during the training phase. Which practice is the most effective technical control to verify training data integrity in an MLOps pipeline?

Accepted Answer

Implementing 'Data Lineage' with checksum validation. Data lineage and cryptographic hashing ensure that the training dataset has not been altered maliciously.

Answer

Increasing model quantization

Answer

Reducing the number of training epochs

Answer

Enabling 'Multi-Factor Authentication' on the training cluster

Question 10

To prevent unauthorized access to an Amazon Bedrock model, which AWS service is the primary point of control for enforcing 'Least Privilege' at the API layer?

Accepted Answer

AWS IAM. AWS IAM (Identity and Access Management) is the primary tool for defining permissions for Bedrock API operations.

Answer

AWS Glue

Answer

AWS CloudFront

Answer

Amazon S3

Question 11

When utilizing the 'Adversarial Robustness Toolbox' (ART) to defend against evasion attacks, which attack type is specifically mitigated by using 'Adversarial Training'?

Accepted Answer

Evasion attacks. Adversarial training involves augmenting the training set with adversarial examples, hardening the model against evasion (inference-time) attacks.

Answer

Model Inversion attacks

Answer

Model Theft attacks

Answer

Supply Chain attacks

Question 12

Which TWO of the following are primary security controls for preventing 'Model Inversion' attacks where an adversary attempts to reconstruct training data from model outputs?

Accepted Answer

Implementing 'Differential Privacy' during training. Differential privacy and output perturbation are the standard methods to reduce the probability of sensitive data leakage through inference queries.

Answer

Setting 'API Rate Limits'

Answer

Enabling 'VPC Service Controls'

Answer

Using 'Hardware Security Modules' (HSM)

Question 13

You are reviewing logs for an AI service. You notice a high frequency of '429 Too Many Requests' errors. Which security control is likely being triggered to protect the model from potential automated extraction or DoS?

Accepted Answer

API Rate Limiting. Rate limiting is a standard control to prevent resource exhaustion and model scraping.

Answer

Model Drift Monitoring

Answer

Data Encryption at Rest

Answer

Input Sanitization

Question 14

Which security principle is violated when an AI model is allowed to access external data sources without sufficient authorization checks on the data itself?

Accepted Answer

Broken Access Control. Unauthorized data access via AI agents often results from 'Broken Access Control' where the AI inherits privileges it should not possess.

Answer

Model Poisoning

Answer

Model Inversion

Answer

Evasion Attack

Question 15

Your organization uses Azure OpenAI. You need to ensure that no sensitive data (e.g., credit card numbers) is sent to the model. Which specific service feature should be enabled to intercept and mask this data?

Accepted Answer

Azure OpenAI Content Safety. Azure OpenAI Content Safety provides text filtering and PII masking.

Answer

Azure Active Directory

Answer

Azure Monitor

Answer

Azure Key Vault

Question 16

When evaluating AI security, 'Model Drift' refers to performance degradation over time. Which tool would a practitioner use to detect drift in a production environment?

Accepted Answer

Amazon SageMaker Model Monitor. Model monitoring tools track statistical properties of features and predictions to identify drift.

Answer

AWS Config

Answer

Amazon GuardDuty

Answer

AWS CloudTrail

Question 17

When securing an AI-driven chatbot, which THREE of the following are recommended practices to mitigate 'Prompt Injection' vulnerabilities?

Accepted Answer

Perform input sanitization on all user-submitted text. The primary defenses involve rigorous input validation, context separation between instructions and data, and using specialized output filters.

Answer

Store model training data in plaintext

Answer

Increase the model's temperature parameter

Question 18

In the context of AI governance, what is the primary role of a 'Model Card'?

Accepted Answer

Provide transparency on model capabilities and limitations. A Model Card is a documentation artifact that provides transparency into the model's intended use, limitations, and training data.

Answer

Grant IAM access to the model file

Answer

Enable encryption for the model weights

Answer

Automatically patch the model software

Question 19

You are implementing 'Red Teaming' for a Large Language Model (LLM). What is the main objective of this exercise?

Accepted Answer

Proactively identify safety and security vulnerabilities. AI Red Teaming identifies vulnerabilities by actively attempting to exploit the model through adversarial prompts.

Answer

Train the model on new, proprietary datasets

Answer

Reduce the cost of API inference

Answer

Increase the accuracy of the model's predictions

Question 20

Which technique is best suited to defend against 'Model Extraction' where an adversary queries a public API to build a clone of your proprietary model?

Accepted Answer

Implement output perturbation. Adding noise (differential privacy) to the outputs or rate-limiting makes the model's decision boundaries difficult to map precisely.

Answer

Enable data augmentation

Answer

Use a larger model architecture

Answer

Rotate API tokens every 30 days