AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 11)

Questions 51–54

1. 51expert · hard

   A security team is scanning a large-scale AI system that includes multiple models, a shared data lake, and a Kubernetes cluster hosting inference services. The team has a limited time window and must prioritize scanning activities. Which scanning approach is most efficient for identifying AI-specific vulnerabilities across the system?

   • APerform a comprehensive scan of the data lake for poisoned samples.
   • BScan the Kubernetes cluster for misconfigurations and known CVEs.
   • CPrioritize scanning the inference endpoints and models, then sample the data lake for poisoning indicators.
   • DRun a standard web vulnerability scanner on all services and skip model-specific tests.
   Reveal Solution Discussion Select an answer first
2. 52foundation · easy

   Which of the following is considered an AI-specific attack surface?

   • AThe backup tape rotation schedule
   • BThe firewall rules for the data center
   • CThe company's email phishing filters
   • DThe inference endpoint that accepts user inputs
   Reveal Solution Discussion Select an answer first
4. 53expert · hard

   A team is fuzzing a reinforcement learning (RL) agent that controls a robotic arm. The team wants to find states that cause the agent to take unsafe actions. They have access to the agent's policy network and a simulator. Which fuzzing strategy is most effective for this scenario?

   • AMutate the reward function to see if the agent changes its behavior.
   • BRandomly sample states from the simulator and record any unsafe actions.
   • COnly test the states that the agent encounters during normal operation.
   • DUse coverage-guided fuzzing on the policy network's activations to find novel states that lead to unsafe actions.
   Reveal Solution Discussion Select an answer first
5. 54application · medium

   A security team wants to fuzz a convolutional neural network (CNN) that classifies medical images. They need a tool that can generate adversarial examples and measure coverage of the model's internal layers. Which type of tool should they select?

   • AA fuzzer that only performs random pixel mutations without any coverage feedback.
   • BA specialized AI fuzzing framework such as TensorFuzz or DeepHunter, which supports neuron coverage and mutation of input tensors.
   • CA standard web application scanner that can send malformed multipart form data to the image upload endpoint.
   • DA general-purpose network fuzzer like AFL, configured to send random images to the model's API.
   Reveal Solution Discussion Select an answer first