Situtations in where AI underperforms during an evaluation to appear safer and lses capable than it truly is.