AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A security auditor is comparing traditional software vulnerability scanning with AI vulnerability scanning for a new project. The auditor wants to ensure that the AI-specific risks are covered. Which statement best describes the scope of AI vulnerability scanning?

   • AIt is identical to traditional scanning but with more aggressive fuzzing.
   • BIt focuses only on the model's accuracy metrics and ignores the surrounding infrastructure.
   • CIt only scans the training data for malicious samples.
   • DIt includes the model, data, and inference endpoints, in addition to the underlying software.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A vulnerability scan reveals that a model is vulnerable to both data poisoning and membership inference. The team has a limited budget and must choose a single mitigation that addresses both issues. Which mitigation is most appropriate?

   • AImplement differential privacy during training.
   • BRetrain the model with a clean dataset.
   • CUse model ensembling to improve robustness.
   • DAdd rate limiting to the inference API.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A company is deploying a large language model (LLM) that uses a vector database for retrieval-augmented generation (RAG). The security team is mapping the attack surface before scanning. Which component should be included as a unique AI-specific attack surface?

   • AThe TLS certificate of the web server hosting the LLM API.
   • BThe vector database's authentication mechanism.
   • CThe prompt injection surface through the LLM's input interface.
   • DThe operating system patches on the GPU servers.
   Reveal Solution Discussion Select an answer first
5. 19expert · hard

   A security team is planning to fuzz a reinforcement learning (RL) agent that controls a robotic arm. They want to find inputs (states) that cause the agent to take unsafe actions. Which fuzzing approach is most appropriate for this AI system?

   • AFuzz the reward function by modifying the environment's reward signals.
   • BGenerate random state vectors without any coverage guidance.
   • CUse a traditional network fuzzer to send random packets to the robot's control interface.
   • DUse a coverage-guided fuzzer that measures neuron coverage of the policy network and mutates state vectors.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Which scanning technique is most appropriate for detecting data poisoning vulnerabilities in an AI system?

   • AAnalyzing the training dataset for anomalies or maliciously injected samples
   • BChecking the model's API rate limits
   • CRunning a network port scanner against the inference endpoint
   • DPerforming a static code analysis of the model's source code
   Reveal Solution Discussion Select an answer first