“The sycophantic behavior exhibited by GPT-4o was a flaw not caught during the evaluation phase, highlighting the limitations of current empirical AI testing methods.”