AI and Machine Learning Fundamentals from an Offensive Security Perspective COASP Practice Questions (Page 5)

Questions 21–25

1. 21foundation · easy

   Which of the following is a component of the AI attack surface that is NOT typically present in traditional software systems?

   • AThe operating system kernel
   • BThe network firewall
   • CThe model's training data pipeline
   • DThe web application front-end
   Reveal Solution Discussion Select an answer first
2. 22application · easy

   A security researcher is analyzing a machine learning model that predicts whether a job applicant will be a good fit. The researcher discovers that the model's predictions are heavily influenced by the applicant's zip code, which correlates with race. This is an example of which type of AI security concern?

   • AModel extraction — the model's decision boundary can be stolen.
   • BData poisoning — the training data was corrupted to introduce bias.
   • CEvasion attack — the model can be fooled by changing the zip code.
   • DBias and fairness — the model is using a proxy for a protected attribute, leading to discriminatory outcomes.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   During an offensive AI assessment, a penetration tester has successfully performed a poisoning attack on a model's training data. According to the standard hacking methodology, which phase does this attack represent?

   • APrivilege escalation
   • BCovering tracks
   • CReconnaissance
   • DExploitation
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   Which adversarial attack involves an attacker querying a machine learning model to reconstruct its decision boundaries or training data?

   • AModel extraction
   • BMembership inference
   • CPoisoning
   • DEvasion
   Reveal Solution Discussion Select an answer first
6. 25expert · hard

   A penetration tester is conducting an assessment of an AI-based intrusion detection system (IDS). The tester has limited time and must choose between two approaches: (1) performing a poisoning attack on the training data, which requires access to the data pipeline, or (2) performing an evasion attack on the live model, which requires crafting adversarial network traffic. The tester has already gained access to the data pipeline but has not yet interacted with the live model. Which approach is more aligned with the exploitation phase of the hacking methodology?

   • APerform reconnaissance on the live model first to gather information
   • BPerform both attacks simultaneously to maximize impact
   • CPerform the poisoning attack, as it directly compromises the model's integrity
   • DPerform the evasion attack, as it is more stealthy and does not require data access
   Reveal Solution Discussion Select an answer first