Data and Training Pipeline Attacks COASP Practice Questions (Page 7)

Questions 31–35

1. 31foundation · easy

   In the context of agentic AI, what is a potential consequence of poisoning the fine-tuning data?

   • AThe agent will lose all access to its external APIs
   • BThe agent may adopt biased or malicious behaviors that align with the attacker's objectives
   • CThe agent will automatically shut down to prevent further damage
   • DThe agent's hardware will overheat due to increased computational load
   Reveal Solution Discussion Select an answer first
2. 32foundation · easy

   How does a backdoor attack affect model behavior at inference time?

   • AThe model misclassifies inputs that contain the trigger pattern, while behaving normally on clean inputs
   • BThe model always outputs random predictions regardless of input
   • CThe model crashes whenever it encounters any input
   • DThe model's accuracy improves on all inputs
   Reveal Solution Discussion Select an answer first
4. 33foundation · easy

   How can a data pipeline attack affect the trustworthiness of an AI system in production?

   • ABy reducing the system's latency
   • BBy causing the system to make errors that erode user confidence and damage the organization's reputation
   • CBy making the system's code open source
   • DBy increasing the system's accuracy on all inputs
   Reveal Solution Discussion Select an answer first
5. 34foundation · easy

   How can data poisoning affect an agentic AI system that uses reinforcement learning?

   • ABy intercepting the agent's actions at inference time and changing them
   • BBy corrupting the reward signals or the environment data used to train the agent, causing it to learn harmful policies
   • CBy deleting the agent's memory of past experiences
   • DBy modifying the agent's neural network architecture during training
   Reveal Solution Discussion Select an answer first
6. 35application · medium

   A hospital uses an AI system to triage radiology images. An attacker corrupts a large portion of the training images by adding random noise and incorrect labels, causing the model to produce unreliable predictions for all patients. The hospital must shut down the system until it is retrained. Which type of poisoning attack and impact does this describe?

   • AIntegrity poisoning, because the attack causes targeted misclassification of specific patients.
   • BBackdoor poisoning, because the attacker embedded a trigger in the corrupted images.
   • CLabel flipping, because the attacker only changed labels without altering image content.
   • DAvailability poisoning, because the attack degrades the model's overall performance and forces it offline.
   Reveal Solution Discussion Select an answer first