[8note]

it seems pretty likely that openai was asking the model to do something bad, and the model did something different thats also bad

if the operator wants something bad, an aligned model should execute on it