AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A security team is reviewing a past incident where an attacker accessed a model's confidence scores through a public API and used them to reconstruct a subset of the training data. Which MITRE ATLAS technique best describes this attack, and which attack surface was primarily exploited?

   • AEvasion via the preprocessing layer
   • BDenial of Service via the inference endpoint
   • CML Model Inference via the model's API response
   • DData Poisoning via the training pipeline
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   Which MITRE ATLAS tactic would you use to categorize an attack that aims to steal the trained model's parameters?

   • AExfiltration
   • BImpact
   • CPersistence
   • DInitial Access
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A financial firm uses a deep learning model for credit scoring. An attacker gains access to the model's training data and inserts a small number of carefully crafted records that cause the model to approve loans for high-risk applicants. Which AI-specific threat does this scenario illustrate, and which attack surface was compromised?

   • AModel inversion; the model's parameters
   • BPrompt injection; the model's input interface
   • CAdversarial examples; the inference API
   • DData poisoning; the training dataset
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A security architect is designing a threat model for a large-scale ML platform that hosts multiple models for different business units. The platform has a shared data lake, a model registry, and a common inference gateway. The architect must prioritize defenses against the most likely attack paths. Which attack path is most critical to address first, given that the platform is internet-facing?

   • AAn attacker exploits a vulnerability in the inference gateway to gain access to the underlying model-serving infrastructure.
   • BAn attacker poisons the shared data lake, affecting multiple models trained on that data.
   • CAn attacker uses the inference gateway to craft adversarial inputs to cause misclassification in a single model.
   • DAn attacker gains write access to the model registry and replaces a model with a malicious one.
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A security analyst is categorizing an attack where an adversary used a publicly exposed Jupyter notebook to modify the model's training code, causing the model to misclassify certain inputs. Using MITRE ATLAS, which tactic and technique combination is most appropriate?

   • AInitial Access via Exploit Public-Facing Application; Impact via Model Poisoning
   • BDiscovery via Model Inference; Persistence via Valid Accounts
   • CExfiltration via Exfiltration Over C2 Channel; Impact via Cost to Organization
   • DEvasion via Adversarial Inputs; Collection via Data from Information Repositories
   Reveal Solution Discussion Select an answer first