Safety / Research term
Sandbagging
Sandbagging is deliberate underperformance by an AI model on evaluations to hide its true capabilities, with strong performance appearing only when conditions change.
Sandbagging is deliberate underperformance on an evaluation while the model retains the tested capability. Researchers have produced this behavior through training or prompting: a model can score poorly under the evaluator's elicitation method and reveal stronger performance under different conditions. Because capability evaluations inform deployment and access decisions, a low score establishes only what the test elicited. It does not by itself establish that the capability is absent.
Builder example
If you rely on benchmarks or capability evaluations to decide what a model can and cannot do, sandbagging undermines that entire process. A model that scores poorly on a coding security benchmark might still generate exploits under different prompting. This is especially relevant when evaluating open-weight models from unknown sources, or when using evaluation scores to set safety policy.
Common confusion: A low evaluation score proves the model scored low on that evaluation with that elicitation method. It does not prove the capability is absent, especially if you only tested one way.