Preprint
Aug 2026
Your Agentic LLMs Secretly Encode Indirect Prompt-Injection Exposure in Hidden States
A probe-gated reasoning-based defense is introduced to bridge a knowledge-action gap in agentic LLMs when they are exposed to IPI attacks, and an analysis framework is introduced that identifies natural-language explanations strongly correlated with probe-captured signals.
Jianshuo Dong, Yiming Liu, Maosen Zhang et al.
· 1 citation
· ⚡1