Design Responsible AI, Security, Governance, Risk Management, and Compliance AB-100 Practice Questions (Page 9)

Questions 41–45

1. 41expert · hard

   A media company is deploying an AI agent that generates news summaries. The agent uses a large language model and is accessible via a public web chat. The company has received reports of users manipulating the agent to produce offensive content. The architect must implement a mitigation that balances user experience and safety. Which approach is most effective?

   • AImplement content filtering on the model's outputs, add input validation to block known malicious prompts, and use Azure AI Content Safety to moderate responses.
   • BRemove the public web chat and require users to access the agent through a VPN.
   • CDisable the agent's ability to generate summaries and instead provide a list of pre-written articles.
   • DUse a smaller language model that is less capable of generating offensive content.
   Reveal Solution Discussion Select an answer first
2. 42application · medium

   A research institute fine-tunes a model on proprietary scientific data. The institute wants to ensure that only the research team can access the fine-tuning dataset, and that any access is logged. What should you implement?

   • AStore the dataset in Azure Blob Storage with a shared access signature (SAS) that never expires.
   • BEncrypt the dataset with a customer-managed key and store the key in the same storage account.
   • CPlace the dataset in a public container and rely on the obscurity of the URL.
   • DUse Azure role-based access control (RBAC) to grant the research team read access to the dataset, and enable diagnostic logging for the storage account.
   Reveal Solution Discussion Select an answer first
4. 43foundation · easy

   Which element is typically included in an audit trail for AI model changes?

   • AThe color scheme of the model's UI
   • BThe number of training epochs
   • CThe model's inference speed
   • DTimestamp of the change and the user who made it
   Reveal Solution Discussion Select an answer first
5. 44application · medium

   A defense contractor is deploying an AI model that classifies satellite imagery. The model is hosted in Azure and accessed by military analysts. The contractor requires that the model be protected from adversarial attacks that could cause misclassification. What should the architect implement?

   • AExpose the model API to the public internet to allow easy access for analysts in the field.
   • BDisable logging to prevent attackers from seeing how the model is used.
   • CUse Azure AI Content Safety to filter inputs, implement rate limiting, and require multi-factor authentication (MFA) for all API access.
   • DUse a shared API key that is distributed to all analysts via email.
   Reveal Solution Discussion Select an answer first
6. 45application · medium

   A technology company hosts a fine-tuned model in Azure Machine Learning. The model is used by an internal application. The security team wants to ensure that only authorized applications can invoke the model and that all invocations are logged. What should you configure?

   • AUse Azure Active Directory (now Microsoft Entra ID) authentication and assign the application a service principal with access to the model endpoint.
   • BEnable Azure Key Vault to store the model weights.
   • CStore the model in a private container registry and share the registry key with all developers.
   • DMake the model endpoint public but obscure the URL.
   Reveal Solution Discussion Select an answer first