Parameter Manipulation COASP Practice Questions (Page 1)

Questions 1–5

1. 1expert · hard

   A tech company's recommendation engine uses a 'diversity_weight' parameter that is passed as a query string in a public API. The parameter is not validated, and an attacker can set it to a negative value, causing the model to return identical recommendations for all users. The company wants to fix this with minimal impact on legitimate API clients. Which approach is best?

   • ARemove the parameter from the API and hardcode the diversity weight.
   • BLog all requests with negative values and ignore them.
   • CReject requests with negative values and return an error message.
   • DClamp the parameter to a valid range (e.g., 0 to 1) before passing it to the model.
   Reveal Solution Discussion Select an answer first
2. 2expert · hard

   A company uses an AI model to detect network intrusions. The model's 'alert_threshold' parameter is set through a web interface that is protected by a username and password. An attacker uses a phishing attack to steal an administrator's credentials and changes the threshold to 1.0. The security team wants to mitigate this risk. Which option is the most effective?

   • AImplement multi-factor authentication (MFA) for the web interface.
   • BAdd a CAPTCHA to the web interface to prevent automated changes.
   • CLog all changes to the threshold and alert on any change.
   • DUse a more robust model that is less sensitive to threshold changes.
   Reveal Solution Discussion Select an answer first
4. 3expert · hard

   A social media company uses a content ranking model that has a 'controversy_penalty' parameter. An attacker discovers they can set this parameter to a negative value through an API, causing the model to boost controversial content. The company's trust and safety team wants to understand the potential impact. Which impact is most concerning?

   • AExposure of user data through the model's outputs.
   • BLoss of model availability because the model crashes.
   • CIncreased user engagement due to more interesting content.
   • DDegradation of model integrity, leading to promotion of harmful content.
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   An e-commerce company uses a recommendation model that reads a 'popularity_weight' hyperparameter from a YAML configuration file. A penetration test reveals that an attacker who gains write access to the server's filesystem can modify this value to zero, causing the model to ignore popularity signals and recommend irrelevant products. The company cannot eliminate filesystem write access because the deployment process requires it. Which mitigation would be most effective to detect this tampering?

   • AAdd a checksum of the configuration file to the application's startup logs.
   • BRemove the popularity_weight parameter from the configuration and hardcode it in the model code.
   • CStore the configuration file in a read-only location and enforce file integrity monitoring (FIM) on it.
   • DEncrypt the configuration file so that only the application can read it.
   Reveal Solution Discussion Select an answer first
6. 5expert · hard

   A company deploys an AI model for autonomous vehicle navigation. The model's 'safety_margin' parameter is stored in a configuration file on the vehicle. The security team wants to prevent tampering, but the engineering team needs to update the parameter over-the-air (OTA) for legitimate improvements. Which solution best meets both requirements?

   • AEncrypt the configuration file so that attackers cannot read it.
   • BUse a more robust model that is less sensitive to changes in the safety margin.
   • CUse a signed OTA update mechanism that verifies the authenticity of the configuration file before applying it.
   • DStore the configuration file in a tamper-evident hardware module that only allows updates through a physical connection.
   Reveal Solution Discussion Select an answer first