Shotgun Jailbreaking Breakdown hackbynight · Follow 10 min read · Just now 5h0tgun J41lbr34k1ng Br34kd0wn J̴̡̢̡̧̧̨̛̛̺̼̫̝̲̦̪͔̮̭͔͙͕͓̱̲͖͓̹̞͉̤͖̬̟͙̟̳̦̯͚̺͔͍̲̓͐͒̾͐͂̈́̑̑̊̓̔͂̈́̂͌̈́̊͐̌̓̊͒͌̉̈͜͜͝͝A̸̧̧̛̰͎͉͖̗̰̩̥̰͎̺̫͍̙̘͖͖̳̤̲̯͔̟̬͖̫̳̫̦̩͍͍̪̘̩͚̳̤̤̟̭̹̙̳̙̜̝͌̀̓̈́͂͋̿̈́̃̒̂͊̈̓̉̃͑̉̊̈̈́͋̉̃̊́̉͛͛̏͋́̐̍̅͐̓̄̀̎̅͘͘̚̕͝͠͠ͅĮ̶̨̛̮̞̣͚̼̲̰̻̮̪̫̳̩̱̠̦̗̺̩̆̓̋̀̓̔͐̍͗̆̄́̂̐̎̉͂̔̿͂͘̚͝͝͠Ľ̴̹̺̬̂̒̈́́̒͒̋̓̀̍͒̊͌̂̑͆̂̌̂̐̾̑̅̉̀͂̈́̊͋͆̑̒̄͒͒̕͘̚͝͝͠͝B̴̡̼̱̞̀̎́̎́̂̈́͒̇́̑̾̀͛̽͐͊̋̍́̊͋̄͂͋͐̇̋̈̉̕͘͝R̵̡̧̨̧̢̢̛̹̹̙͕̞͔̻̞̯͔̙͉̤̰͉̗̩̟̗̹̟͈̯͙̳̠͖̹̭͕͈̺̦̫̔͋Ȅ̷̢̡̡̢̡̡̛̩̹̹͚̲̥̬͕̞̮͙̹̠͎̭͚̫̗̫̲͉͉̱̖͓͔̥̯̦͍͈͎̫̹͓̻̬͉͙̟̦̬̓̃̎̄͐̔̆́̏̄̈́̽̓̇̂̂̃́̐̏̒͑̓̈̂̔̓̍̀̆͐̕͜͝͝͠͝͝ͅÄ̸̡̨̢̧̞̟̱̥͓͔͇͎̖̫̰͖͙̱̖̞̖̮̱̗̰̖̩͚̙̦͎͉̙̯͙̖͚̞̺́̔͛̿͐͑̌͆͜ͅĶ̷̢̡̡̢̪͉̱͈̘͔̹̘̼͙͕̟̜͔̠̬̠̤̘̞͙̤̲̳̞̯̗͙̓̅͂́̌̅̄͆̃͊̔̓̔̈́̀͗͗̍̉̽͒̋̚͜͝͠͠͠S̵̢̛̱̟͇̣̗̺̘̣̳̖̙̦̘̰̥̀͗͒̓͂́͗͌̓̉̾̐̿͗͛̔͌̄̆̽̄͌̀̈̉̓̀͂̊̌̈́̀̚͝͝͝ FOR ALL FLAGSHIP AI MODELS Just months after exposing the sophisticated “many-shot jailbreaking” technique, Anthropic has explored another critical vulnerability in Large Language Models — one that’s both simpler and potentially more concerning. This new method, dubbed “shotgun jailbreaking” or “best-of-n jailbreaking,” represents a further evolution in how we understand AI system vulnerabilities. In my last blog post covering many-shot jailbreaking, I explored how attackers could methodically “teach” LLMs to bypass their safety measures through carefully crafted examples. Many-Shot Jailbreaking Breakdown. Anthropic, creators of the Claude… | by hackbynight | Medium That technique required understanding model behavior and crafting sophisticated prompt sequences. What makes this new discovery particularly interesting is its brutal simplicity: […]
Original web page at blog.hackbynight.nl
