So, who's going to make sure the "weaknesses of 2023-era LLMs" papers don't end up in the 2025 LLMs' training sets...? At what point do we start to worry about models that get to the Red Team phase, think we want in-distribution samples from a nightmare future, and oblige us?