NEW in the outline effective 1 August 2026. AI and ML workloads hosted in cloud environments, the security properties and risks specific to model training and inference, and how AI services change the shared responsibility picture. This subdomain does not exist in pre-2026 prep material.
AI and ML in cloud environmentsmodel training and inference securityAI shared responsibilityAI governanceNIST AI Risk Management Framework
Practice question for this objective
lock_openFree sampleCloud Concepts, Architecture and Designhard
A cloud provider offers a managed large language model that a customer consumes purely through an inference API, sending prompts and receiving completions. Under the AI shared responsibility model, which security obligation belongs to the customer rather than the provider?
AHardening the GPU host operating system and hypervisor that execute the model weights
BPatching the model training pipeline and retraining framework libraries
CApplying differential privacy to the corpus used to pretrain the foundation model
DGoverning what data is placed into prompts and validating the model output before acting on itcheck_circle Correct
In a managed AI inference service the customer owns prompt data governance and output validation while the provider owns the model and infrastructure. The shared responsibility boundary follows control: the consumer controls only the data it submits and how it uses the results, so input classification and output validation are its duties, whereas the training pipeline, weights and host stack are operated and therefore secured by the provider.
Why A is wrong: This is tempting because model inference is compute-intensive and the customer pays for that compute, but in a managed inference service the provider owns the underlying host, hypervisor and accelerators, so this is provider responsibility.
Why B is wrong: This sounds correct because vulnerable libraries are a real risk, but the customer using only an inference API does not operate the training pipeline; the provider builds and maintains the managed model, so this sits with the provider.
Why C is wrong: Privacy of the training corpus is a genuine concern, but the pretraining corpus and its privacy controls belong to the provider that trained the managed model, not to a customer that only calls inference.
Why D is correct: In a consumption model the customer controls the inputs it sends and the use it makes of the outputs, so classifying prompt data, preventing sensitive data leakage into prompts, and validating completions are the customer's responsibility.
Exam traps in Cloud Concepts, Architecture and Design
Answers that look right on this material and are not. Each one is a distractor from a different question in the CCSP bank for this domain.
Enable customer-managed encryption keys for the object store holding the training corpus
errorWhy it is wrong: Customer-managed keys protect data at rest and address key custody, but they do nothing to stop a trained model from leaking individual records through its inference outputs.
Increase the size of the training dataset so the model generalises across more citizen categories
errorWhy it is wrong: More training data may improve accuracy but a larger corpus alone gives no visibility into how decisions are made and no ongoing bias measurement, so it misses the governance ask.
Public cloud, because the infrastructure is provisioned for open use by the general public and shared across unrelated tenants.
errorWhy it is wrong: Public cloud is provisioned for open use by many unrelated tenants; here the infrastructure serves one organisation exclusively, so this does not fit.