Agent Security Is a Systems Problem
arxiv.org
2 threads
This position paper argues that classic systems security design principles are what we need to secure AI agents.
agent security is a very old lesson UNlearned.
data and instructions don't mix. And agents can't separate them