AAISM AI Technologies And Controls Practice Test 4 — 15 Questions

Question 1

Which of the following is a 'Control' for mitigating the risk of 'Prompt Injection'?

Accepted Answer

Sandboxed execution environment for agents. A sandboxed execution environment for agents is the correct control because prompt injection attacks aim to make an LLM-driven agent execute unintended actions (e.g., calling tools, accessing files, or making network requests). Sandboxing constrains what the agent can actually do even if the prompt is manipulated, enforcing least privilege at the execution layer. This is a defense-in-depth control that limits blast radius rather than trying to perfectly filter malicious prompts.

Answer

Using a larger language model

Answer

Providing the agent with full root access

Answer

Disabling output validation

Question 2

What is 'Model Stealing' and why is it considered a security incident?

Accepted Answer

It is the theft of intellectual property via query analysis. Model stealing is the extraction of a model's functionality or intellectual property through repeated query analysis — an attacker probes the API with crafted inputs and uses the input-output pairs to train a surrogate model that approximates the original. It is a security incident because it compromises the owner's IP and can enable further attacks such as adversarial example transfer and evasion of the original model. The economic value of a trained model is directly eroded when it can be cloned via the inference API.

Answer

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

Answer

It is the unauthorized access to the server room

Answer

It is the accidental deletion of model files

Question 3

What is the goal of 'Data Sanitization' before training?

Accepted Answer

Remove PII and malicious data from the training set. Data sanitization is the process of cleaning a training dataset by removing personally identifiable information (PII), toxic content, and malicious or poisoned samples before the data is used to train a model. This reduces privacy risk, regulatory exposure, and the chance that an attacker can manipulate model behavior through data poisoning. The goal is not to change formats or reduce costs but to ensure the training data is safe and compliant.

Answer

Change the file format of the dataset

Answer

Reduce the cloud storage costs

Answer

Increase the number of images in the dataset

Question 4

Which THREE of the following are examples of AI system 'Access Controls'?

Accepted Answer

API Key authentication. API Key authentication (B) is an access control because it requires a caller to present a valid key before the AI service or endpoint will accept the request, authenticating and gating access. Role-Based Access Control (C) is a classic access control mechanism that grants permissions to AI resources and operations based on a user's assigned role, enforcing least privilege. Network Security Groups (D) act as access controls by allowing or denying inbound and outbound traffic to AI workloads based on rules, thereby restricting which sources can reach the system. Disabling logging services (A) is not an access control but rather the removal of an auditing/monitoring capability, and increasing model training epochs (E) is a model-training hyperparameter change that has no bearing on who or what may access the AI system.

Answer

Disabling logging services

Answer

Increasing model training epochs

Question 5

Which tool is best for managing 'Secrets' (like database passwords or API keys) used by your AI training pipeline?

Accepted Answer

HashiCorp Vault. HashiCorp Vault is purpose-built for secrets management, offering dynamic secrets, encryption as a service, leasing and revocation, and detailed audit logs. It integrates with identity providers and cloud IAM to issue short-lived credentials for databases, API keys, and AI pipeline components, so no static secret ever lives in code or config. For an AI training pipeline that touches sensitive datasets and third-party model APIs, Vault's centralized policy engine and audit trail directly support governance and least-privilege requirements.

Answer

Hardcoded variables in the training script

Answer

A public Github repository

Answer

An Excel spreadsheet on a shared drive

Question 6

In the context of 'Model Poisoning', what is a 'Backdoor' attack?

Accepted Answer

Embedded trigger causing specific model behavior. A backdoor attack in model poisoning involves an adversary implanting a hidden trigger pattern into the training data so the model behaves normally on clean inputs but produces attacker-chosen outputs whenever the trigger is present. The trigger can be a rare token, a specific pixel pattern, or a syntactic phrase. This is distinct from general poisoning because the malicious behaviour is conditional and stealthy, evading standard accuracy benchmarks.

Answer

Deleting the model's training documentation

Answer

Increasing the model's complexity

Answer

Bypassing authentication servers

Question 7

Which THREE of the following are common steps in an 'AI Incident Response Plan'?

Accepted Answer

Containment of the affected AI model. Option E is correct because identification and classification of the AI incident is the essential first phase of any incident response plan, establishing what happened, its severity, and the appropriate response track before other actions are taken. Option A is correct because containment of the affected AI model limits blast radius by isolating, quarantining, or disabling the model (e.g., revoking API keys, taking the endpoint offline, or rolling back to a known-good version) so the incident does not spread or cause further harm. Option C is correct because eradication of the threat source removes the root cause — such as poisoned training data, a compromised model artifact, a malicious prompt-injection vector, or an exploited vulnerability — so the incident cannot recur after recovery. Option B is not a response step; increasing inference speed is a performance-tuning activity unrelated to incident handling. Option D is not a standard response step; wholesale hardware replacement is a disproportionate, costly action that is not part of a typical AI incident response plan.

Answer

Increasing the model's inference speed

Answer

Replacing all hardware in the data center

Question 8

What is the primary role of 'Data Encryption at Rest' for AI datasets?

Accepted Answer

Protect data from unauthorized access on storage media. Data encryption at rest protects AI datasets by rendering stored data unreadable without the correct cryptographic keys, preventing unauthorized parties who gain access to storage media (stolen disks, misconfigured buckets, compromised backups) from reading the data. It is a confidentiality control, not a performance, detection, or patching mechanism.

Answer

Increase the training speed of the model

Answer

Detect potential poisoning attacks

Answer

Automatically patch model software

Question 9

You are configuring a 'Firewall' for an AI API. Which rule should you prioritize to prevent unauthorized exfiltration of model training data?

Accepted Answer

Configure strict egress (outbound) traffic rules. Strict egress (outbound) traffic rules control what data can leave the AI environment, directly preventing unauthorized exfiltration of training data or model artifacts. By defaulting to deny and allowing only approved destinations, ports, and protocols, you block covert channels an attacker might use to send stolen data out. This is the highest-priority firewall control for data exfiltration prevention.

Answer

Disable all firewall logging

Answer

Increase the number of API endpoints

Answer

Allow all outbound traffic to the internet

Question 10

What is the 'Membership Inference' attack in machine learning?

Accepted Answer

Determining if a specific record was in the training set. Membership inference determines whether a specific record was part of the model's training set by observing the model's outputs (e.g., confidence scores) for that record. It is a privacy attack because it can reveal sensitive information — for example, that a particular patient's data was used to train a medical model, which itself leaks health status. The attack exploits the tendency of models to be more confident on training data than on unseen data.

Answer

Brute forcing the user password

Answer

Adding noise to the output

Answer

Stealing the model's entire weights file

Question 11

You detect an unusual spike in API calls that are clearly non-human. What is the most effective way to address this without completely blocking the service?

Accepted Answer

Implement a bot detection challenge (e.g., CAPTCHA). A CAPTCHA or similar bot detection challenge is the most effective way to mitigate non-human API traffic without blocking legitimate users. It presents a challenge that is easy for humans but difficult for automated scripts, thereby filtering out bots while allowing genuine users to continue accessing the service. This approach is targeted and minimally disruptive, aligning with the goal of not completely blocking the service.

Answer

Increase the model latency

Answer

Publicly release the model weights

Answer

Shutdown the entire API service

Question 12

Why is it important to perform 'Model Validation' before production deployment?

Accepted Answer

To verify performance and safety against benchmarks. Model validation before production verifies that the model meets defined performance, fairness, robustness, and safety benchmarks on representative data, ensuring it behaves as intended in real-world conditions. It is a governance gate that catches degradation, bias, or unsafe behavior before customers or critical processes are affected. Removing users, encrypting weights, and reducing GPU costs are unrelated to validation's purpose.

Answer

To automatically remove users from the system

Answer

To encrypt the model weights

Answer

To reduce the cost of GPU hardware

Question 13

Which THREE of the following are technical controls to prevent 'Model Extraction' attacks?

Accepted Answer

API Rate Limiting. API Rate Limiting (A) is correct because throttling the number and frequency of inference queries per client makes it economically and practically infeasible to harvest enough input-output pairs to reconstruct a model's decision boundary. Output perturbation (B) is correct because injecting calibrated noise into returned predictions degrades the fidelity of the surrogate model an attacker can train, while preserving utility for legitimate users. Reducing output detail/granularity (D) is correct because returning only coarse labels or rounded confidence scores instead of full probability vectors removes the rich gradient-like information attackers rely on for high-fidelity extraction. Increasing the model learning rate (C) is not a security control at all—it is a training hyperparameter that affects convergence and could even destabilize the model—and providing all intermediate model outputs (E) is the opposite of a defense, since exposing internal representations gives attackers far more signal to clone the model.

Answer

Increasing the model learning rate

Answer

Providing all intermediate model outputs

Question 14

What is the role of 'Homomorphic Encryption' in secure AI? (Theoretical concept)

Accepted Answer

Allows computation on encrypted data. Homomorphic encryption allows computations to be performed directly on encrypted data without decrypting it first, producing an encrypted result that, when decrypted, matches the result of the same computation on the plaintext. In secure AI, this enables model inference or training on sensitive data while the data remains encrypted end-to-end, protecting privacy even from the compute provider. This is the defining property of homomorphic encryption.

Answer

Detects malicious prompts

Answer

Manages API user identities

Answer

Compresses the model size

Question 15

An AI administrator is using Azure AI Content Safety to detect jailbreak attempts in prompt engineering. Which configuration must be enabled to proactively block prompt injection attacks?

Accepted Answer

Configure Prompt Shield. Prompt Shield is the Azure AI Content Safety feature specifically designed to detect and block prompt injection and jailbreak attempts, including both user prompt attacks and indirect attacks embedded in documents. Enabling it applies a dedicated detection model that flags adversarial prompts before they reach the underlying LLM. This is the only option that directly addresses jailbreak and prompt injection mitigation.

Answer

Apply Customer Managed Keys (CMK)

Answer

Enable Protected Material Detection

Answer

Implement Jailbreak Mitigation via Language Detection