AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A company is deploying a large language model (LLM) for internal document summarization. The security team is identifying attack surfaces. Which component is most likely to be exploited by a prompt injection attack?

   • AThe training data repository
   • BThe model's inference API
   • CThe GPU acceleration hardware
   • DThe model's weight files
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A red-team analyst is documenting an attack where an adversary used a publicly available API to query a sentiment-analysis model thousands of times, and from the confidence scores, reconstructed a rough approximation of the model's decision boundary. The analyst wants to classify this activity using MITRE ATLAS. Which tactic and technique pair is most appropriate?

   • AInitial Access – Exploit Public-Facing Application
   • BML Model Inference – Infer Training Data Membership
   • CML Model Inference – Extract ML Model
   • DReconnaissance – Search for Victim's Publicly Available Research Materials
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A security analyst is categorizing an attack where an attacker used a compromised CI/CD pipeline to replace a model file in the deployment bucket. The replacement model behaves differently on certain inputs. Which MITRE ATLAS technique and attack surface pair is most accurate?

   • AExfiltration; the deployment bucket
   • BData Poisoning; the training dataset
   • CML Model Evasion; the inference endpoint
   • DML Model Poisoning; the model registry
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A red team is documenting an attack where they used a phishing email to obtain credentials for a cloud ML platform, then accessed the training data and modified it. Which MITRE ATLAS tactic and technique pair best describes the credential access and the data modification?

   • AEvasion via Adversarial Inputs; Collection via Data from Information Repositories
   • BInitial Access via Phishing; Impact via Data Poisoning
   • CExfiltration via Exfiltration Over C2 Channel; Impact via Cost to Organization
   • DDiscovery via Model Inference; Persistence via Valid Accounts
   Reveal Solution Discussion Select an answer first
6. 10foundation · easy

   Which of the following best describes the 'deployment' component of an AI system's attack surface?

   • AThe accuracy metric used for evaluation
   • BThe physical servers hosting the AI model
   • CThe format of the training data files
   • DThe algorithm used to train the model
   Reveal Solution Discussion Select an answer first