Scholars found by gathering as little as a hundred examples of question-answer pairs for illicit advice or hate speech, they could undo the careful “alignment” meant to establish guardrails around ...