Capability overhang

A situation in which a model already has hidden, untapped capabilities exceeding its common uses, revealed only through better prompting, fine-tuning, or tool access. A system's real capabilities can outrun expectations, making risk assessment harder. A concept from AI safety discourse.