1 paper · 1 filter
Dominik Seip, Bartol Markovinović, Matthias Hein
The safety of LLM-based agents currently rests on alignment training and safety classifiers, neither of which can rule out that a sophisticated indirect prompt injection (IPI) goes…