AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 8)

Questions 36–40

1. 36application · medium

   A security team is analyzing a known attack where an attacker submitted a specially crafted image to a facial recognition system, causing it to identify the wrong person. The team wants to document the attack surface and technique. Which combination is correct?

   • ATraining dataset; Data Poisoning
   • BPreprocessing pipeline; Denial of Service
   • CModel repository; Model Theft
   • DInference API; Adversarial Inputs
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   A security engineer is mapping the attack surface of a customer-facing image classification service. The service uses a pre-trained model hosted on a GPU VM, with images uploaded via a web API and stored in blob storage before inference. The engineer wants to identify where an attacker could most plausibly inject a malicious input that would be processed by the model without needing to compromise the VM or storage. According to MITRE ATLAS, which attack surface should be prioritized?

   • AThe model's training pipeline, because poisoned training data would alter the model's behavior for all users.
   • BThe web API endpoint, because it is the only externally reachable component that accepts untrusted data for inference.
   • CThe GPU VM's network interface, because the attacker could send crafted packets to the inference service.
   • DThe blob storage container, because an attacker could upload a malicious image if the storage key is leaked.
   Reveal Solution Discussion Select an answer first
4. 38application · medium

   A healthcare startup deploys an on-premises image classification model for detecting skin lesions. The model is trained on a public dataset and served via a REST API. The security team is performing an attack surface review. Which component of the AI system presents the highest-risk attack surface for a data poisoning attack?

   • AThe REST API endpoint that accepts inference requests
   • BThe public training dataset used to fine-tune the model
   • CThe on-premises GPU servers hosting the model
   • DThe model's final classification threshold setting
   Reveal Solution Discussion Select an answer first
5. 39application · medium

   A red-team analyst is categorizing an attack where an adversary used a public API to query a translation model with carefully crafted sentences to extract the model's internal vocabulary and sentence structure. The analyst wants to map this to MITRE ATLAS. Which technique is most appropriate?

   • AML Attack Staging – Craft Adversarial Data
   • BExfiltration – Exfiltration via ML Inference API
   • CML Model Inference – Infer Training Data Membership
   • DML Model Inference – Extract ML Model
   Reveal Solution Discussion Select an answer first
6. 40expert · hard

   A security architect is designing a defense for a medical imaging model. The model is hosted on-premises and accessed via a web application. The architect is concerned about both model inversion and adversarial examples. Which attack surface should be prioritized to mitigate both threats?

   • AThe model's version control repository
   • BThe GPU hardware
   • CThe web application's API responses
   • DThe training data storage
   Reveal Solution Discussion Select an answer first