AI Attacks PT0-003 Practice Questions (Page 3)

Questions 11–15

1. 11expert · hard

   A company's image-recognition model is used to identify defective products on a manufacturing line. An attacker gains access to the model's training pipeline and injects a small number of images of defective products labeled as non-defective. The model is retrained and now fails to detect some defects. The company wants to detect this attack. Which defense is the most effective?

   • AImplement anomaly detection on the training data
   • BUse a larger model
   • CImplement rate limiting on the model's API
   • DUse cryptographic hashing on the model file
   Reveal Solution Discussion Select an answer first
2. 12foundation · easy

   Which of the following is a method to protect AI models from manipulation?

   • AUsing cryptographic hashes to verify the integrity of model files before deployment
   • BUsing a larger batch size during training
   • CIncreasing the model's temperature parameter during inference
   • DEncrypting the model's output
   Reveal Solution Discussion Select an answer first
4. 13expert · hard

   A company's AI assistant is integrated with a database of customer records. The assistant is used by customer service representatives to look up customer information. An attacker gains access to the assistant's API and sends a prompt that tricks it into revealing all customer records. The company wants to prevent this without affecting the assistant's ability to look up individual customer records. Which defense is the most effective?

   • AImplement output filtering to block responses that contain more than one customer record
   • BAdd a disclaimer to the assistant's responses
   • CDisable the assistant's ability to access the internet
   • DUse a separate model for customer service queries
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A company uses a machine-learning model to detect fraudulent transactions. An attacker gains access to the model's storage and replaces the model file with a malicious version that approves all transactions. Which defense would be most effective in detecting this?

   • AImplement rate limiting on the model's API
   • BDisable the model's internet access
   • CUse a larger model
   • DUse cryptographic hashing to verify model integrity
   Reveal Solution Discussion Select an answer first
6. 15foundation · easy

   Which of the following is a vulnerability that makes an AI system susceptible to prompt injection?

   • AThe AI model is hosted on a cloud platform
   • BThe AI model uses a large training dataset
   • CThe AI model is trained on publicly available data
   • DThe AI model does not distinguish between instructions and untrusted data in the input
   Reveal Solution Discussion Select an answer first