1 paper
Georgios Syros, Blake Bullwinkel, Alina Oprea +4
Existing large language model (LLM) backdoor attacks operate primarily in chat settings, where the risk is bounded by free-form text outputs. AI agents remove that bound by perform…