205 questions total · 3pages · All types, answers revealed
Page 3 of 3
What is the primary function of 'IAM' in an AI-driven cloud environment?
IAM provides the foundation for Least Privilege.
Why this answer
IAM (Identity and Access Management) manages who can perform which actions on which resources (e.g., who can invoke a model).
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?
Security Groups enforce network-level access control for container instances.
Why this answer
Security Groups provide stateful packet filtering for ECS containers, restricting who can communicate with the model endpoint.
When a model exhibits 'Hallucinations', what is the best strategy to improve its groundedness and security?
RAG provides a source of truth for the model to reference.
Why this answer
RAG (Retrieval-Augmented Generation) forces the model to base its answers on specific, provided documents, reducing hallucination risks.
When securing an AI-driven chatbot, which THREE of the following are recommended practices to mitigate 'Prompt Injection' vulnerabilities?
Sanitization removes characters or patterns used in injection attacks.
Why this answer
The primary defenses involve rigorous input validation, context separation between instructions and data, and using specialized output filters.
You are securing a PyTorch model and want to implement differential privacy to protect the training data. Which library is specifically designed to work with PyTorch for this purpose?
Opacus is the standard privacy library for PyTorch.
Why this answer
Opacus is the official PyTorch library for training models with differential privacy.
Which THREE activities help manage 'AI Lifecycle' risk?
Lifecycle management starting point.
Why this answer
Design review, monitoring, and regular decommissioning/sunset reviews cover the start, middle, and end of the lifecycle.
Which metric should be used to track 'Data Quality' for AI training?
Standard quality metrics.
Why this answer
Data completeness, accuracy, and representativeness are key metrics for assessing the quality of training sets.
A researcher is using 'Adversarial Training' to defend against evasion. How exactly does this work?
This is the core definition of adversarial training.
Why this answer
Adversarial training includes adversarial examples in the training set, allowing the model to learn to classify them correctly.
What is the purpose of 'Logging' in an AI system?
Comprehensive logging is required for security forensics.
Why this answer
Logging provides the audit trail necessary to investigate security incidents and understand how the system reached a specific output.
What is the primary risk associated with 'Data Poisoning' in an AI governance program?
Direct impact of poisoned data.
Why this answer
Poisoned data compromises the model's integrity and reliability, leading to biased or malicious outputs.
What is the goal of 'Data Sanitization' before training?
This protects against data leakage and poisoning.
Why this answer
Data sanitization removes sensitive information (PII/PHI) and malicious patterns, ensuring that the model does not learn or leak private info.
Which THREE of the following are potential indicators of a 'Model Poisoning' attack occurring during the training phase?
Corrupted data often prevents smooth convergence.
Why this answer
Poisoning often results in degraded accuracy for specific targets (backdoors) or broader performance degradation that is detectable through validation testing.
What is the primary risk of using 'Pre-trained models' from public hubs without verification?
This is the classic supply chain risk for unverified models.
Why this answer
Public hubs are susceptible to supply chain attacks where models may contain backdoors or malicious weights.
When securing a large language model (LLM) using the OWASP Top 10 for LLMs, which control is most effective against 'Insecure Output Handling'?
Sanitizing model output before downstream execution prevents cross-site scripting or code injection.
Why this answer
Insecure Output Handling occurs when a model output is interpreted by a downstream system without sanitization. Output validation (canonicalization) is the primary defense.
A firm is drafting a policy for 'AI Use in Marketing'. Which element is most important to include to ensure alignment with existing data protection regulations?
Transparency and disclosure are critical regulatory and governance requirements for AI-generated consumer content.
Why this answer
Ensuring the AI system complies with transparency and privacy requirements (like GDPR) is foundational to AI governance policy.
Which THREE of the following are common steps in an 'AI Incident Response Plan'?
Stopping the impact is essential.
Why this answer
Incident response involves identification (detection), containment (isolation), and eradication/recovery, followed by post-incident learning.
What is the goal of an 'AI Security Awareness Program'?
Fundamental for reducing human error and insider risk.
Why this answer
To ensure all staff understand the risks of AI, including prompt injection, data leakage, and ethical pitfalls.
Which THREE documents are vital for an AI audit trail?
Essential for security/compliance audit.
Why this answer
Model version, training set, and validation report are the three pillars of a reconstructible audit trail.
Which THREE of the following are effective ways to secure the 'ML Pipeline' (MLOps) against supply chain attacks?
Isolation prevents lateral movement.
Why this answer
Securing the pipeline involves code signing, scanning dependencies for vulnerabilities, and using isolated container environments.
Which cryptographic tool is used to ensure the authenticity of a model file stored in an S3 bucket?
Signatures provide integrity and authenticity verification.
Why this answer
Digital signatures (using tools like Cosign or AWS KMS) verify that the model has not been tampered with since being signed.
What is the 'Membership Inference' attack in machine learning?
This exploits the model's tendency to memorize training samples.
Why this answer
This attack determines whether a specific data point was included in the training set, potentially compromising privacy if the data is sensitive.
Who is ultimately responsible for the outcomes of an AI system deployed by a business unit?
Accountability is linked to business ownership.
Why this answer
The Business Owner or designated executive holds accountability for the system's performance, risk, and compliance.
A developer is implementing 'Differential Privacy' on a training dataset. What is the primary trade-off they are managing?
The privacy budget directly controls the noise level, impacting model accuracy.
Why this answer
Differential privacy adds noise to training data to preserve privacy, which inherently reduces the utility or accuracy of the model.
Which THREE items should be included in an 'AI Vendor Assessment'?
This is critical for understanding the product's risks.
Why this answer
Vendor assessments must verify security practices, compliance, and model transparency.
Which THREE technical configurations help mitigate the risk of 'Model Exfiltration'?
Controlling file access is the most direct defense against exfiltration.
Why this answer
Model exfiltration can be hindered by restricting access to the model, monitoring egress traffic, and obfuscating model responses.
When evaluating AI 'Bias', which concept must be governed first?
Without definition, bias cannot be measured or managed.
Why this answer
Defining fairness metrics is the foundational step before measurement or mitigation can occur.
How should an organization classify AI systems for governance?
Classification should be based on risk potential.
Why this answer
Risk-based classification (e.g., low, medium, high) allows the organization to allocate resources proportional to potential impact.
What is the primary risk associated with 'Shadow AI' in an organization?
Lack of oversight is the core risk of shadow IT.
Why this answer
Shadow AI refers to the use of AI tools without IT or security oversight, leading to data leaks and compliance violations.
A company is developing an internal LLM-based tool. Which activity is required for 'Accountability'?
Provides the basis for accountability.
Why this answer
Documenting the training data, model parameters, and design decisions is necessary to hold the developers and owners accountable.
Which TWO strategies are recommended for 'Human-in-the-Loop' governance?
Enables effective HITL.
Why this answer
Standardizing the process and defining the specific decision points are key governance strategies.
An AI Security analyst is configuring Google Cloud Vertex AI Model Garden to ensure compliance with organization policies. Which IAM permission is strictly required to prevent unauthorized users from deploying a base model to a public endpoint?
This permission controls the creation of endpoints, which is where public/private visibility is defined.
Why this answer
The roles/aiplatform.user role allows usage, but 'aiplatform.endpoints.create' or specific restrictions on the Endpoint resource are needed to prevent public exposure.
When reviewing an AI architecture, what does the 'Attack Surface' refer to?
Reducing this surface is a core security objective.
Why this answer
The attack surface includes all points where an adversary can enter or extract data from the system, including API endpoints and training interfaces.
To monitor for AI security incidents, you are integrating logs from an AI gateway into a SIEM. Which specific field in the request log is most critical for detecting 'Model Denial of Wallet' attacks?
Token count directly correlates to billing and consumption.
Why this answer
Denial of Wallet attacks involve high-cost, high-token-count requests. Monitoring the 'usage.total_tokens' or 'billing_tokens' field is critical for identifying this behavior.
When designing a secure AI development lifecycle (SDLC), which of the following is an essential governance activity during the 'Model Deployment' phase?
Documentation via Model Cards is a standard governance practice to ensure accountability and transparency in production.
Why this answer
A 'Model Card' is a documentation standard that captures risk parameters and intended use, essential for post-deployment governance.
Which THREE of the following are examples of AI system 'Access Controls'?
API keys authorize the caller to use the service.
Why this answer
Access controls include authentication (IAM), network restrictions (VPC/Security Groups), and fine-grained permissions for API operations.
Which TWO of the following are considered 'AI Supply Chain' risks?
Data poisoning is a critical supply chain risk.
Why this answer
AI supply chain risks include dependencies on external data and untrusted pre-trained models.
Which TWO of the following are essential components of an 'AI Security Policy' in an enterprise?
Validation is a required security activity.
Why this answer
A robust policy must define acceptable use cases and mandate security testing (like red teaming) for all AI assets.
To mitigate adversarial evasion attacks against a vision model, you decide to implement defensive distillation on your TensorFlow model. What is the primary operational trade-off you must evaluate?
Distillation smooths the gradient surface, which makes it harder for attackers but can impact general prediction accuracy.
Why this answer
Defensive distillation increases the model's robustness against small adversarial perturbations but can lead to reduced accuracy on clean data or increased training complexity.
Which TWO factors contribute to 'Model Obsolescence' risk?
The existence of better tech makes old models obsolete.
Why this answer
Obsolescence occurs when models become ineffective due to changing data distributions (drift) or superior new models.
Which THREE steps are needed for 'AI Risk Assessment'?
Threat modeling identifies potential attack vectors.
Why this answer
Assessment requires identifying the AI components, evaluating threats, and determining the business impact.
When dealing with 'Adversarial Evasion', why is the 'black-box' nature of an API a risk?
Black-box APIs provide the information needed for query-based evasion.
Why this answer
In a black-box attack, an attacker can craft perturbations by repeatedly querying the API and analyzing the outputs to build a substitute model.
In the context of 'Model Poisoning', what is a 'Backdoor' attack?
The trigger activates the malicious function.
Why this answer
A backdoor is a trigger (a specific pattern) embedded into the model during training so that it behaves normally until the trigger is present, at which point it performs a malicious action.
Which TWO steps are recommended to secure access to an AI model registry?
Limits model access to authorized users and service accounts.
Why this answer
Access control for registries should follow the principle of least privilege and use identity federation.
Which THREE security controls are effective against 'Data Leakage' in AI?
Prevents sensitive data from entering prompts.
Why this answer
DLP, access control, and output filtering are the standard triad for preventing sensitive data exposure via AI.
What is the most effective control to prevent 'Model Scraping' from an exposed AI prediction API?
Rate limiting restricts the volume of requests a scraper can perform.
Why this answer
Rate limiting and request authentication are the standard controls to prevent automated extraction of model knowledge.
Which THREE strategies are effective for defending against model poisoning in an ML pipeline?
Helps verify if the model has learned the desired patterns.
Why this answer
Defending against poisoning requires cleaning the training data and using robust learning algorithms.
Which of the following describes 'Model Integrity' risk?
This is a direct integrity risk.
Why this answer
Model integrity risk refers to the potential for the model to be altered or behave in an unauthorized way, such as through poisoning or unauthorized weight updates.
You are implementing 'AI Red Teaming'. What is the most effective way to test against prompt injection?
This is the core methodology of red teaming for LLMs.
Why this answer
AI Red Teaming involves systematically probing the model with adversarial prompts to evaluate its response under stress.
Which TWO actions should be taken when implementing an AI security monitoring strategy in a production environment?
Detects potential adversarial attacks in real-time.
Why this answer
Effective monitoring requires both tracking system health and analyzing input/output traffic for malicious patterns.
What is 'AI Shadowing' or 'Shadow AI' in an enterprise setting?
This is the definition of shadow AI and poses significant governance risk.
Why this answer
Shadow AI refers to the use of AI tools and models by employees without the knowledge or approval of the IT/Security department.
Which THREE metrics should be monitored for 'Model Performance' and security?
Critical indicator of performance.
Why this answer
Drift, latency, and error rates are foundational metrics for system health and security monitoring.
Which activity is part of 'Ongoing AI Monitoring'?
Core task of operational AI governance.
Why this answer
Continuous performance review ensures that the AI system remains within its defined risk appetite and operational parameters.
Which of the following best describes 'Model Inversion' as a security risk?
Reconstruction of training samples is the definition of model inversion.
Why this answer
Model inversion involves exploiting model outputs to reconstruct the training data, potentially revealing sensitive information.
What is 'AI Risk Appetite'?
Definition of risk appetite.
Why this answer
It is the level of risk the organization is willing to accept in pursuit of its AI-driven business objectives.
An organization is implementing 'Red Teaming' for AI. What is the main governance objective?
Core objective of red teaming.
Why this answer
Red teaming identifies vulnerabilities and threat vectors to proactively improve the resilience of the AI system.
Page 3 of 3
Practice AAISM by domain
Target a specific domain to shore up weak areas.
See all domains with question counts →