cialdini-authoritypretends to be an expert or official so the model defers to the claimed authority.foot-in-doorstarts with a tiny harmless request, then escalates once the model has said yes.fear-appealwarns of bad outcomes to scare the model into complying.guilt-manipulationmakes the model feel responsible so it bends its rules.bandwagon-pressureclaims everyone else already does it, so the model should too.
Attack Strategies
Persuasion: Attack Strategy
Attacks that use human-style social pressure to talk the AI into breaking its rules. No code tricks, just psychology.
What it is: Attacks that use human-style social pressure to talk the AI into breaking its rules. No code tricks, just psychology.
How the attacks work: The attacker leans on classic influence tactics. Claim authority, build up small agreements, create fake urgency, or guilt the model. Many run over several turns, getting the AI to commit a little at a time until it gives in.
Real examples from the framework: