How Adversarial Environments Mislead Agentic AI?

arXiv:2604.18874v1 Announce Type: new Abstract: Toolintegrated agents are deployed on the premise that external tools ground their outputs in reality. Yet this very reliance creates a critical attack surface.