AI RESEARCH
Trust No Tool: Evaluating and Defending LLM Agents under Untrusted Tool Feedback
arXiv CS.CL
•
ArXi:2605.17453v1 Announce Type: cross Tool-using LLM agents increasingly rely on external tools to make consequential decisions, yet most existing agent-security benchmarks and defenses implicitly assume that tool feedback is trustworthy once a tool has been selected. We study a different failure mode, cognitive poisoning, in which a malicious tool behaves plausibly during exploration, accumulates trust through benign-looking feedback, and becomes harmful only when hidden state conditions align with the final executable action.