Contamination is the presence of benchmark content, or closely related material, in a model’s training data or memory. In AI security and evaluation, this matters because it can make a system look more capable than it really is: the model may be recalling prior exposure instead of solving the task fresh. That weakens the value of scores, especially on static tests that are reused widely.
In real-world defenses, contamination is treated like a measurement and integrity problem. Teams look for duplicate or near-duplicate items, hidden prompt leakage, and training data overlap so they can separate memorization from general reasoning. Attackers can exploit contamination indirectly by seeding public data with benchmark-like content, or by overfitting systems to known test patterns, which can distort evaluations and hide weaknesses that would matter in deployment.