Attention-based Classification for Indirect Prompt Injection

Amelia Kawasaki, Stefan Lee

Indirect prompt injection (IPI) attacks, in which adversarial instructions are embedded in external data processed by a language model, pose a significant threat to the integrity o…