AI Security Glossary
Key terms and concepts in AI security testing
Prompt Injection
User input that manipulates model behavior.
Model Hallucination
Model invents non-factual output.
Data Poisoning
Training data manipulated to bias outcomes.
SBOM/MBOM
Software/Model Bill of Materials, for supply chain transparency.
Adversarial Example
Input designed to cause model misclassification.
Model Stealing
Extracting model parameters through queries.
Membership Inference
Determining if data was in training set.
Differential Privacy
Technique to protect individual data points.
Red Teaming
Simulating adversary attacks to test defenses.
Jailbreaking
Bypassing model safety restrictions.