The teach Single-turn safety checks miss multi-turn manipulation. Jailbreaks work by exploiting conversational momentum. The attacker changes roles, justifications, or context until the model treats the unsafe request as newly legitimate. The last turn looks bad, but the earlier turns did the shaping work. That means your guardrails must survive: reframing: “this is for education” or “this is fictional” decomposition: splitting one harmful task into smaller requests persistence: repeated asks designed to exhaust the refusal pattern Good defenses remember intent across the conversation and escalate when the pattern, not just the sentence, looks adversarial.
Read the full lesson
Sign up free: one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in