Model Vulnerabilities

This category covers attacks and tests targeting the AI model itself, beyond the prompt interface. AI Red Teamers investigate inherent weaknesses in the model's architecture, training data artifacts, or prediction mechanisms, such as susceptibility to data extraction, poisoning, or adversarial manipulation.

Learn more from the following resources:

• @article@AI Security Risks Uncovered: What You Must Know in 2025
• @article@Weaknesses in Modern AI
• @report@AI and ML Vulnerabilities (CNAS Report)