Reposted by Dan Davies
I continue to feel like this thing where they RL models on impossibly huge amounts of exploit and infosec stuff and then are surprised when the models use those techniques when not specifically asked to is a little bit something or other.
