CSA Trusted AI Safety Expert (TAISE)

Which of the following best describes a Large Language Model (LLM)?

• AA deep neural network pre-trained on vast text corpora to predict and generate human-like…
• BA rule-based expert system trained on a fixed knowledge base
• CA supervised classifier that assigns labels to discrete input categories
• DA reinforcement learning agent optimized for game-playing environments

In the context of AI, what does the term 'hallucination' refer to?

• AThe model generating outputs that are factually incorrect or entirely fabricated with app…
• BA visual distortion artifact caused by adversarial image perturbations
• CA mode of training where the model imagines counterfactual scenarios to improve robustness
• DAn overfit condition in which training loss is extremely low but validation loss is high

What is the primary purpose of the tokenization step in a Generative AI pipeline?

• ATo convert raw text into numerical representations (tokens) the model can process
• BTo encrypt model weights before inference
• CTo filter harmful content before it reaches the end user
• DTo compress training data for storage efficiency

Retrieval-Augmented Generation (RAG) primarily addresses which LLM limitation?

• ASlow inference speed due to large model size
• BLack of mathematical reasoning capability
• CKnowledge cutoff and inability to access real-time or proprietary information
• DVulnerability to adversarial prompt injection

Which transformer component is most directly responsible for capturing relationships between all tokens in an input sequence simultaneously?

• AFeed-forward neural network layer
• BPositional encoding
• CMulti-head self-attention mechanism
• DLayer normalization

Fine-tuning a pre-trained LLM on domain-specific data is best categorized as which type of machine learning approach?

• ATransfer learning adapted to a target domain
• BUnsupervised pre-training from scratch
• CReinforcement learning from human feedback
• DZero-shot inference without any additional training

A healthcare organization deploys an LLM to draft patient-facing summaries. The primary ethical concern with this use case is:

• APotential for the model to generate incorrect medical information that harms patients
• BHigh GPU power consumption during inference
• CThe open-source nature of the model weights
• DExcessive model latency causing poor user experience

Which of the following is an example of an AI use case that carries a HIGH risk of generating deepfakes?

• AUsing LLMs to summarize internal quarterly earnings reports
• BApplying computer vision to inspect manufacturing defect rates
• CUsing text-to-video generative models to synthesize realistic video of public figures
• DDeploying a recommendation engine for e-commerce product suggestions

Algorithmic bias in an AI hiring tool is most likely to cause which type of harm?

• ASystematic discrimination against protected groups in candidate selection
• BData exfiltration from the HR database
• CIncreased cloud infrastructure costs due to model complexity
• DUnauthorized access to model training weights

LIME (Local Interpretable Model-agnostic Explanations) is used to:

• AGenerate locally faithful explanations for individual model predictions by approximating…
• BEncrypt model predictions before sending them to end users
• CDetect adversarial examples in model inputs
• DReduce the dimensionality of training data to speed up convergence

Which principle states that AI systems must be designed so their actions and decision-making processes can be understood and traced by humans?

• AAutonomy
• BRobustness
• CExplainability
• DScalability

In the AI model lifecycle, 'data poisoning' most commonly occurs during which phase?

• AModel deployment and serving
• BPost-deployment monitoring and retraining
• CTraining data collection and ingestion
• DUser-facing inference

Model drift occurs when a deployed model's performance degrades over time. Which type of drift is caused by changes in the statistical properties of input features?

• AData drift (covariate shift)
• BConcept drift
• CLabel drift
• DModel poisoning

Which MLSecOps practice helps maintain AI model integrity by continuously checking production predictions against a baseline quality threshold?

• AContinuous model monitoring with statistical drift detection
• BBlue-green model deployment
• CDifferential privacy during training
• DFederated learning across distributed nodes

The NIST AI Risk Management Framework (AI RMF) organizes its guidance into four core functions. Which of the following is NOT one of them?

• AGovern
• BMap
• CMeasure
• DRespond

ISO/IEC 42001 is best described as:

• AThe first certifiable international standard for an AI Management System (AIMS)
• BA technical standard specifying neural network architecture requirements
• CA regulation governing AI liability across EU member states
• DA cryptographic standard for protecting machine learning model weights

MITRE ATLAS (Adversarial Threat Landscape for AI Systems) is primarily used to:

• ACatalog and communicate adversarial machine learning tactics, techniques, and procedures…
• BDefine regulatory compliance requirements for AI in financial services
• CMeasure model bias across demographic subgroups
• DCertify cloud providers for AI workload hosting

The CSA AI Controls Matrix (AICM) contains how many control objectives across how many domains?

• A100 control objectives across 10 domains
• B156 control objectives across 12 domains
• C243 control objectives across 18 domains
• D300 control objectives across 20 domains

Which NIST AI RMF function is responsible for establishing organizational policies, roles, and accountability structures for AI risk?

• AMAP
• BMEASURE
• CMANAGE
• DGOVERN

A prompt injection attack against an LLM-powered application involves:

• AInserting malicious instructions into user input or retrieved content to override the mod…
• BInjecting SQL commands into the model's database connection
• CPhysically tampering with GPU firmware to alter inference outputs
• DFlooding the model API with requests to cause a denial of service

Model inversion attacks aim to:

• ARecover private training data by querying the model and analyzing its outputs
• BFlip the sign of model gradients to slow down training
• CReverse-engineer the model's source code from its API responses
• DBypass content filters by encoding malicious inputs in Base64

Which defense technique adds carefully calibrated random noise to training data or model outputs to protect individual privacy?

• ADifferential privacy
• BHomomorphic encryption
• CFederated learning
• DModel watermarking

An adversarial example is best defined as:

• AA carefully crafted input with imperceptible perturbations that causes a model to make a…
• BA deliberately mislabeled training sample used to mislead the model during training
• CAn example drawn from a different domain than the training distribution
• DA synthetic data point generated by a GAN to augment training

Which cloud deployment model is most commonly used for training large-scale foundation models due to elastic compute and managed GPU clusters?

• APrivate cloud on-premises data center
• BCommunity cloud shared by regulated-industry members
• CPublic cloud (hyperscaler IaaS/PaaS)
• DEdge computing nodes at inference endpoints

Zero Trust applied to AI workloads means that:

• AEvery request to an AI service must be continuously verified regardless of source network…
• BAI models are never granted network access to prevent data leakage
• COnly zero-shot models that require no training data are permitted
• DAI systems are deployed only in air-gapped environments with no external connectivity

Which security control is most effective at protecting sensitive data used to train AI models stored in a cloud object store?

• AEncrypting data at rest with customer-managed keys (CMK) and enforcing least-privilege IA…
• BEnabling public access to simplify data ingestion by training clusters
• CStoring training data only in memory-resident caches to avoid persistent storage
• DCompressing training data with a lossless algorithm to obscure its structure

What is the primary purpose of a Model Bill of Materials (MBOM) or AI Bill of Materials (AI BOM)?

• ATo provide a transparent inventory of components, datasets, dependencies, and licenses us…
• BTo document the financial cost of training and deploying an AI model
• CTo list the names of human annotators who labeled the training data
• DTo record inference latency benchmarks under various load conditions

Which technique allows multiple organizations to collaboratively train a shared AI model without sharing their raw data?

• AFederated learning
• BTransfer learning
• CActive learning
• DCurriculum learning

Under GDPR, using personal data to train an AI model without a valid legal basis constitutes:

• AAn unlawful processing activity that can result in fines up to €20 million or 4% of globa…
• BA minor policy violation correctable by a written warning
• CA breach of the EU AI Act's high-risk AI requirements
• DA violation of ISO 42001 Annex A controls only

Data minimization in AI systems means:

• ACollecting and retaining only the personal data strictly necessary for the specified AI p…
• BUsing the smallest possible model architecture to reduce compute costs
• CApplying lossy compression to reduce the size of training datasets
• DLimiting the number of output tokens generated per inference call

Training data anonymization through k-anonymity ensures that:

• AEach record is indistinguishable from at least k-1 other records on quasi-identifier attr…
• BAll records are encrypted with a 256-bit AES key
• CThe dataset contains at least k synthetic records for every real record
• DData is partitioned into k equal shards before federated training

Which emerging AI threat involves attackers injecting hidden instructions into content that an AI agent will process (e.g., a web page or email), causing the agent to perform unintended actions?

• AIndirect prompt injection
• BModel extraction
• CGradient leakage
• DMembership inference

Jailbreaking an LLM refers to:

• AUsing specially crafted prompts to bypass the model's safety guardrails and elicit prohib…
• BBreaking encryption on model weights stored in a cloud bucket
• CRemoving the model from its container image to run it on bare-metal infrastructure
• DExtracting an LLM's training data by brute-forcing its API

Which threat involves an attacker querying an AI model API extensively to reconstruct a functional replica of the model?

• AModel inversion
• BMembership inference
• CModel extraction (model stealing)
• DBackdoor attack

A 'Trojan' or backdoor attack on a machine learning model is characterized by:

• AEmbedding a hidden trigger pattern during training so the model behaves normally on clean…
• BUsing trojaned hyperparameters to slow down the training convergence
• CInserting a remote-access Trojan into the model's deployment server
• DAdding a hidden layer to the model architecture to reduce inference accuracy

In the context of AI supply chain security, which risk is most directly associated with using a third-party pre-trained model from a public repository?

• AThe model may be slower than a custom-trained model
• BThe model will not support fine-tuning on proprietary data
• CThe model may contain embedded backdoors, poisoned weights, or malicious code introduced…
• DThe model requires proprietary hardware accelerators not available in standard cloud envi…

Which of the following best describes 'model collapse' in the context of training future AI models on AI-generated content?

• AA catastrophic hardware failure that destroys model weights during training
• BThe intentional pruning of model layers to reduce size below a minimum threshold
• CA phenomenon where models trained on AI-generated data progressively lose diversity and a…
• DA distributed training failure caused by network partitioning between GPU nodes

Which AI safety concept describes the property that an AI system should behave as intended across diverse, edge-case, and adversarial conditions?

• AReliability
• BScalability
• CRobustness
• DInterpretability

Output filtering and content moderation guardrails for LLMs are best applied:

• AOnly at training time using RLHF alignment
• BOnly at the model layer using system prompts
• CAt multiple layers: input filtering, system prompt, model alignment, and output filtering
• DExclusively at the network perimeter using a WAF

Which of the following is NOT a recommended practice for securing an AI model API endpoint?

• AImplementing rate limiting and throttling to prevent API abuse and DoS
• BLogging all API requests and responses for audit and anomaly detection
• CPublishing the full system prompt publicly to enable community review
• DEnforcing authentication and authorization on every API call

The NIST AI RMF Generative AI Profile (NIST AI 600-1) introduces which AI-specific risk not covered in the original AI RMF 1.0?

• AData quality issues in traditional ML pipelines
• BBias in classification model outputs
• CConfabulation (hallucination) as a named AI risk category unique to generative AI
• DSecurity vulnerabilities in on-premises ML infrastructure

An AI ethics review board within an organization serves primarily to:

• AAccelerate model deployment by pre-approving all AI projects in bulk
• BManage GPU cluster capacity and cloud spending for AI workloads
• CProvide oversight, evaluate ethical risks, and ensure AI deployments align with organizat…
• DConduct red-team penetration tests against deployed AI models

The EU AI Act classifies AI systems used for biometric identification of individuals in public spaces by law enforcement as:

• AMinimal risk, requiring only a code of conduct
• BLimited risk, requiring transparency notices
• CHigh risk, subject to strict conformity assessments
• DProhibited AI practices, with narrow exceptions for specific national security purposes

Which metric is most appropriate for evaluating whether an AI model treats different demographic subgroups equally in a binary classification task?

• AF1 score averaged across all samples
• BRoot mean squared error (RMSE)
• CDemographic parity (statistical parity difference)
• DPerplexity score on a held-out test set

ISO/IEC 42001 Annex A controls specifically address which AI-related concern?

• ACryptographic key management for model weight protection
• BNetwork segmentation requirements for GPU clusters
• CAI system impact assessment, responsible use objectives, data quality management, and hum…
• DOpen-source license compliance for third-party AI libraries

Which of the following best characterizes the relationship between AI safety and AI security?

• AThey are synonymous terms for the same discipline
• BAI security is a subset of AI safety restricted to data encryption
• CAI safety focuses on preventing unintended harmful outcomes; AI security focuses on defen…
• DAI safety applies only to autonomous vehicles; AI security applies to all other AI systems