citation-manipulationgets the model to produce fake or wrong citations that look real.false-premisehides a false claim inside the question so the answer accepts it.fabrication-promptingasks for details the model does not know, inviting it to make them up.confidence-exploitationpushes the model to sound certain about things it cannot verify.sycophancy-exploitleans on the model’s urge to agree with the user even when the user is wrong.
Attack Strategies
Misinformation: Attack Strategy
Tricks that nudge an AI into stating false things as if they were true. The aim is wrong but confident answers, fake sources, or biased framing.
What it is: Tricks that nudge an AI into stating false things as if they were true. The aim is wrong but confident answers, fake sources, or biased framing.
How the attacks work: The attacker shapes the question so a false answer feels natural. A baked-in wrong assumption, a demand for a citation, or flattery can all push the model to agree instead of correct.
Real examples from the framework: