1 paper
Basavesh Ammanaghatta Shivakumar, Swarn Priya, Peng Gao
LLM agents increasingly read untrusted content, invoke external tools, access private data, and delegate work to other agents. Harm often arises not from a single unsafe action but…