Parameter Manipulation COASP Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A company deploys a chatbot that uses a large language model (LLM) with a 'temperature' parameter that controls response randomness. The chatbot's API accepts a 'temperature' field from the client. An attacker sets 'temperature' to 2.0, causing the chatbot to produce incoherent and potentially harmful responses. Which of the following is the best way to mitigate this while preserving the chatbot's functionality?

   • AAdd a content filter to the chatbot's responses to block harmful outputs.
   • BRemove the 'temperature' parameter from the API entirely and use a fixed value.
   • CUse a different LLM that is not affected by temperature changes.
   • DReject or clamp any 'temperature' value outside the allowed range (e.g., 0.0–1.5) at the API gateway.
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A self-driving car company uses a deep learning model for object detection. The model takes camera images as input and has a 'confidence_threshold' parameter that is set via a configuration file on the vehicle. An attacker with physical access to the vehicle modifies the configuration file to set the threshold to 0.99. What is the most likely effect on the vehicle's behavior?

   • AThe model will only detect objects with very high confidence, potentially missing many real objects and increasing the risk of collisions.
   • BThe model will crash because the threshold is too high.
   • CThe model will operate normally because the threshold is within the valid range.
   • DThe model will detect more objects, including false positives, because it is more sensitive.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   An autonomous vehicle company uses a perception model that takes a 'confidence_threshold' parameter from a configuration file. A researcher demonstrates that lowering this threshold causes the model to detect obstacles that are not present, leading to unnecessary braking. Which statement best describes the relationship between the parameter and the model's behavior?

   • AThe parameter changes the model's training data distribution.
   • BThe parameter modifies the model's neural network architecture.
   • CThe parameter has no effect on the model's output.
   • DThe parameter directly controls the trade-off between precision and recall in the model's detections.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   Which of the following is an example of a parameter manipulation attack surface within an AI model's configuration?

   • AThe model's training dataset
   • BThe model's loss function implementation
   • CModel hyperparameters stored in a configuration file
   • DThe model's output labels
   Reveal Solution Discussion Select an answer first
6. 15application · medium

   A financial services company exposes a fraud-detection API that accepts a 'risk_threshold' parameter in the JSON request body. A security analyst notices that lowering this threshold causes the model to flag far more transactions, and an attacker could set it to zero to force every transaction to be reviewed, creating a denial-of-service condition. The team wants to prevent this without breaking legitimate API clients that need to adjust sensitivity. Which control should they implement?

   • ALog all API requests that include the risk_threshold parameter and review them weekly for anomalies.
   • BRemove the risk_threshold parameter from the API entirely and hardcode the model's threshold in the configuration file.
   • CEncrypt the risk_threshold parameter using the client's public key so only the model can decrypt it.
   • DValidate and clamp the risk_threshold parameter to an allowed range (e.g., 0.1 to 0.9) on the API gateway before it reaches the model.
   Reveal Solution Discussion Select an answer first