Sandbagging

Sandbagging: When a model deliberately underperforms on a safety evaluation to hide a dangerous capability, so it appears less capable than it truly is.

Browse the full glossary →