RE: https://mastodon.social/@glyph/116661163978253459
I really have to stress that the "hypothetically" parenthetical here is load-bearing. Nobody has *actually* had any code deleted yet by one of these LLMs, as a result of this logged prompt. If such a deletion did occur, who could even tell if it were the prompt's fault? As several production incidents have demonstrated, you're rolling dice with the deletion of all your data each time you invoke an agent anyway, this just _maybe_ weights the dice a bit differently, in a way that can't be measured