Aaditya Pai

Real-world documents expose the inadequacy of traditional prompt injection defenses, with PARSE achieving a significant reduction in attack success while preserving utility.

Paraphrasing content before processing slashes domain-camouflaged attack success rates by up to 84%, but effectiveness varies drastically by model and domain.

LLM safety filters are virtually blind to adversarial attacks that blend in with the target domain, dropping detection rates to near zero.