AAISM AI Technologies And Controls Practice Test 2 — 15 Questions

Question 1

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?

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

Accepted Answer

API Rate Limiting. HTTP 429 'Too Many Requests' is the standard status code returned when a client exceeds a defined request rate. API rate limiting is the security control that enforces these thresholds, protecting the model endpoint from abuse such as automated model extraction (stealing the model via repeated queries) or denial-of-service. By capping requests per identity or IP, rate limiting preserves availability and limits data exfiltration.

Answer

Model Drift Monitoring

Answer

Data Encryption at Rest

Answer

Input Sanitization

Question 2

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. Broken Access Control (A) is violated when an AI model accesses external data sources without proper authorization checks on the data itself — the model is effectively bypassing the access control layer that should govern which data it can read. This is the OWASP Top 10 #1 category and applies directly to AI systems that pull from APIs, databases, or file stores without enforcing per-resource authorization. The model should only reach data its service identity is explicitly permitted to access.

Answer

Model Poisoning

Answer

Model Inversion

Answer

Evasion Attack

Question 3

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. Amazon SageMaker Model Monitor is purpose-built to detect data drift, model quality drift, bias drift, and feature attribution drift in production endpoints. It captures inference data, compares it against baselines, and emits CloudWatch alarms when drift exceeds thresholds. This directly addresses the need to detect performance degradation over time in a deployed model.

Answer

AWS Config

Answer

Amazon GuardDuty

Answer

AWS CloudTrail

Question 4

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. Input sanitization (A) is correct because stripping or escaping control characters, delimiters, and instruction-like patterns from user-submitted text reduces the attacker's ability to smuggle directives that override the intended prompt. Using system messages to enforce strict context boundaries (B) is correct because a privileged system prompt that defines the assistant's role, allowed scope, and instruction hierarchy helps the model distinguish trusted developer instructions from untrusted user content. Deploying a refusal layer for unexpected system-level queries (D) is correct because detecting and rejecting requests that attempt to access system prompts, tools, or internal configuration blocks escalation paths typical of prompt injection. Option C is wrong because storing training data in plaintext increases exposure and does nothing to mitigate prompt injection. Option E is wrong because raising the temperature parameter increases randomness and creativity, making outputs less predictable and potentially easier to manipulate, rather than hardening the chatbot against injection.

Answer

Store model training data in plaintext

Answer

Increase the model's temperature parameter

Question 5

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 structured document that discloses a model's intended use, training data, performance metrics, limitations, and ethical considerations, providing transparency to stakeholders. Its primary governance role is to communicate what the model can and cannot do so that consumers, auditors, and risk teams can make informed decisions. This transparency function is the defining purpose of model cards in AI governance frameworks.

Answer

Grant IAM access to the model file

Answer

Enable encryption for the model weights

Answer

Automatically patch the model software

Question 6

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. The main objective of red teaming an LLM is to proactively identify safety and security vulnerabilities before adversaries exploit them. Red teaming involves adversarial testing—crafting prompts designed to bypass guardrails, elicit harmful outputs, or extract sensitive information—to uncover weaknesses that can then be remediated. It is a defensive exercise focused on discovery, not on model improvement or cost reduction.

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 7

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. Output perturbation deliberately adds noise or rounds confidence scores returned by the API, degrading the fidelity of the information an adversary can use to train a clone. Model extraction relies on precise output signals, so perturbing them reduces clone accuracy while preserving acceptable utility for legitimate users. This is a direct, low-cost mitigation at the inference boundary.

Answer

Enable data augmentation

Answer

Use a larger model architecture

Answer

Rotate API tokens every 30 days

Question 8

Which TWO of the following are effective methods for monitoring and detecting AI security incidents in a cloud environment?

Accepted Answer

Ingesting model inference logs into a SIEM. Option C is correct because ingesting model inference logs into a SIEM centralizes AI telemetry, enabling correlation of prompts, responses, and metadata with other security events to detect prompt injection, data exfiltration, or model abuse in near real time. Option E is correct because configuring anomaly detection for API call patterns baselines normal invocation frequency, latency, token usage, and source IPs, then alerts on deviations that indicate credential theft, model extraction, or denial-of-wallet attacks against AI endpoints. Option A is wrong because disabling logging removes the audit trail needed for detection and forensics, and cost reduction is not a security monitoring method. Option B is wrong because standard database backup schedules address data durability and recovery, not real-time detection of AI security incidents. Option D is wrong because hardcoding API credentials in application code exposes secrets and increases compromise risk rather than helping monitor or detect incidents.

Answer

Disabling logging to reduce cloud spend

Answer

Using standard database backup schedules

Answer

Hardcoding API credentials in application code

Question 9

When training an AI model, why is it critical to ensure 'Data Provenance'?

Accepted Answer

To verify the source and integrity of training data. Data provenance establishes a verifiable record of where training data originated, who handled it, and whether it was altered — this is foundational for trustworthiness, regulatory compliance, and detecting poisoned or tampered datasets. Without provenance, you cannot attest to model behavior, reproduce training runs, or defend against supply-chain attacks on the data pipeline.

Answer

To reduce the storage space of datasets

Answer

To automatically remove PII from datasets

Answer

To optimize the training speed of the model

Question 10

You need to ensure that a model deployed on AWS ECS is only accessible by authorized microservices. What is the most appropriate network security control?

Accepted Answer

Use AWS Security Groups. AWS Security Groups act as virtual firewalls at the instance/ENI level and control inbound and outbound traffic based on protocol, port, and source/destination. To ensure a model on ECS is only accessible by authorized microservices, you attach a security group to the ECS tasks that allows traffic only from the security groups of those authorized microservices. This provides least-privilege network segmentation without exposing the service publicly.

Answer

Increase the container memory limit

Answer

Enable AWS WAF for all internal traffic

Answer

Use public routing for all endpoints

Question 11

Which of the following describes the purpose of 'Differential Privacy' in a machine learning system?

Accepted Answer

To prevent the identification of individuals in the training data. Differential Privacy (DP) is a mathematical framework that adds calibrated noise to data or computations so that the inclusion or exclusion of any single individual's data has a negligible effect on the output. In ML, it prevents attackers from determining whether a specific person was in the training set or inferring their attributes, thereby protecting individual privacy while still allowing aggregate statistical learning.

Answer

To enable cross-cloud model training

Answer

To allow for better interpretation of model layers

Answer

To ensure the model is faster at inference time

Question 12

Which THREE of the following are potential indicators of a 'Model Poisoning' attack occurring during the training phase?

Accepted Answer

Inconsistent model convergence during training. Option B is correct because model poisoning corrupts the training data or update process, so the loss curve and parameter updates no longer converge consistently, producing erratic or stalled convergence during training. Option C is correct because poisoned training samples can cause the model to memorize manipulated patterns, yielding high training-set accuracy while validation accuracy degrades due to the corrupted decision boundary. Option E is correct because attackers often implant backdoors or trigger-based behaviors, so the model behaves normally on ordinary inputs but produces attacker-chosen outputs when a specific trigger pattern is present. Option A is not a training-phase indicator, since increased API request latency reflects inference or service performance rather than poisoned training data. Option D is also not indicative, because larger model-weight storage is a resource/cost artifact and does not by itself demonstrate poisoned training behavior.

Answer

Increased latency for API requests

Answer

Increase in storage costs for the model weights

Question 13

What is the primary function of an 'AI Firewall' or 'LLM Gateway' in an enterprise AI architecture?

Accepted Answer

To intercept and filter prompts and responses. An AI Firewall or LLM Gateway sits between users/applications and the underlying model, acting as an inline policy enforcement point that inspects, filters, and sometimes transforms prompts and responses. Its core purpose is to intercept traffic to block prompt injection, jailbreaks, data leakage, toxic content, and policy violations before they reach the model or the end user.

Answer

To optimize the GPU utilization of the model

Answer

To store model weights in a secure archive

Answer

To automate the labeling of training data

Question 14

Which of the following is a 'Human-in-the-Loop' (HITL) control used for AI risk management?

Accepted Answer

Requiring human approval for high-risk model outputs. Human-in-the-Loop (HITL) means a human reviews, approves, or overrides model outputs before they take effect, which is a direct risk control for high-stakes decisions. Requiring human approval for high-risk model outputs ensures accountability and catches errors, bias, or unsafe recommendations before harm occurs. This is a core control in AI risk management frameworks for high-impact use cases.

Answer

Using only open-source libraries

Answer

Deploying the model to a public server

Answer

Automating all model output without review

Question 15

You are managing a model's lifecycle in SageMaker. How do you ensure that only approved model versions are deployed to production?

Accepted Answer

Use 'SageMaker Model Registry' approval workflows. SageMaker Model Registry provides versioned model groups with approval statuses (PendingManualApproval, Approved, Rejected), and deployment pipelines can be gated to only deploy models in the Approved state. This creates an auditable, policy-enforced promotion path from training to production.

Answer

Use manual file copying to the production server

Answer

Delete all non-production model versions

Answer

Set every model version to 'Production' by default