Many-Shot Jailbreaking: How Giant Context Windows Trick Filters
Faits validés
Anthropic Research (April 2024, arXiv:2404.03094). Because modern AIs can process hundreds of thousands of words of context, attackers feed in hundreds of fabricated Q&A dialogues until the model uncritically continues the pattern.
En termes simples
The habituation tactic: Tell a bouncer 99 harmless stories and on the 100th he will nod through what should actually be forbidden.
Pertinence sécurité & conseil pratique
New AI capabilities (larger context windows) automatically create entirely new security holes.