AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 1)

Questions 1–5

1. 1application · medium

   A healthcare organization uses an ML model to predict patient readmission risk. The model is trained on electronic health records (EHR) and deployed in a cloud environment. A researcher discovers that by querying the model with synthetic patient profiles, they can infer whether a specific real patient was in the training dataset. Which AI-specific threat is this?

   • AMembership inference, because the attacker is determining if a specific data record was part of the training set.
   • BAdversarial example, because the attacker is crafting inputs to cause misclassification.
   • CModel extraction, because the attacker is learning the model's decision boundary.
   • DData poisoning, because the attacker is injecting malicious data into the training set.
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A security analyst is reviewing a known incident where attackers inserted malicious code into an open-source library that was later incorporated into a company's ML training pipeline. The code exfiltrated a subset of the training data to an external server. The analyst wants to map this to MITRE ATLAS. Which combination of tactic and technique best describes the exfiltration of the training data?

   • AExfiltration – Exfiltration via Cyber Means
   • BCollection – Collect Data from ML Artifacts
   • CImpact – Cost to Develop ML System
   • DExfiltration – Exfiltration via ML Inference API
   Reveal Solution Discussion Select an answer first
4. 3expert · hard

   A security analyst is evaluating the threat landscape for a generative AI model that produces code. The analyst is concerned that an attacker could craft prompts that cause the model to generate malicious code. Which AI-specific threat and attack surface does this scenario represent?

   • AAdversarial examples; the preprocessing layer
   • BPrompt injection; the model's input interface
   • CData poisoning; the training dataset
   • DModel inversion; the model's parameters
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A bank deploys a fraud-detection model that reviews credit-card transactions. An attacker has learned that the model is retrained weekly on a rolling window of recent transactions. The attacker submits a series of small, legitimate-looking transactions from a set of accounts they control, gradually biasing the model to classify their larger fraudulent transactions as normal. Which AI-specific threat does this scenario best illustrate?

   • AAdversarial examples, because the attacker is crafting inputs that are misclassified at inference time.
   • BData poisoning, because the attacker is injecting malicious samples into the training data to corrupt the model's behavior.
   • CPrompt injection, because the attacker is manipulating the model's input text to override its instructions.
   • DModel inversion, because the attacker is trying to extract the model's training data.
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A security analyst is mapping an attack where an attacker exploited a vulnerability in a model's preprocessing library to execute arbitrary code on the server. Which attack surface and ATLAS technique is most appropriate?

   • AThe model's inference API; ML Model Evasion
   • BThe model's deployment environment; Denial of Service
   • CThe preprocessing pipeline; Exploitation for Client Execution
   • DThe training dataset; Data Poisoning
   Reveal Solution Discussion Select an answer first