The security risk in tool-augmented LLM agents arises from conflating action induction with execution authorization

this risk arises from conflating action induction with execution authorization

http://arxiv.org/abs/2608.27146v1