held-out payload D1 naive regex D2 normalize+broad D3 normalize+strict h1 'new task' reframing - - - h2 'please do not mention' - - - h3 'IMPORTANT' banner - - - h4 German override - - - h5 reversed text - - - h6 'ignore the above' phrase - Y Y h7 role-play jailbreak - - - h8 URL exfil via link - - - TOTAL detected of 8 0 1 1