AAISM Poisoning Practice Questions

A SaaS-based LLM system has risks including prompt injection, data poisoning, and model exfiltration. What is the BEST way to ensure consistent risk treatment?

• AA. Apply control baselines from a recognized industry standard
• BB. Implement an AI threat control matrix mapping threats to controls and assurance
• CC. Focus on post-deployment red teaming
• DD. Rely on vendor audit reports and SLAs

The PRIMARY goal of data poisoning attacks is to:

• AA. compromise the confidentiality of output data from the model
• BB. compromise the confidentiality of model input data
• CC. manipulate the behavior of the model during development
• DD. undermine the integrity of the AI system's outputs

Which of the following strategies is the MOST effective way to protect against AI data poisoning?

• AA. Ensuring the model is trained on diverse data sources
• BB. Increasing model complexity
• CC. Using robust data validation techniques and anomaly detection
• DD. Incorporating more features and data into model training