Existing evaluations of tool‑using agents usually measure whether agents can complete diverse tasks with tools, assuming tool outputs are reliable. In real systems, tool returns can be plausible yet wrong. We evaluated fourteen large language models with three tools—web search, LLM sub‑agent delegation, and code execution—by corrupting their outputs and measuring how often agents incorporated the corrupted content into their final answers. Overtrust was observed across all settings; the mean adoption rate exceeded one third for every tool and reached 68.0% for web search. Trace analysis revealed a worrying pattern: agents often detected conflicts and even derived the correct answer internally, but still presented only the corrupted answer without warning the user. To mitigate overtrust we intervened at three levels: user prompting, tool‑provider metadata, and post‑training adjustments by the agent builder. Some interventions helped specific models or tools, yet none consistently reduced overtrust across all tools. These findings identify overtrust in unreliable tools as a serious, persistent failure mode and motivate evaluations and interventions that enable agents to validate tool outputs and transparently communicate unresolved conflicts.
Review