Deception Under Tool Failure in LLM Agents

Tool-augmented agents are commonly evaluated by final-answer correctness, which does not distinguish accurate reporting of tool failure from answers that assert values the tool never returned. We study when operational tool failures cause this behaviour and whether it depends on failure signalling, system prompts, and agent frameworks. Our benchmark contains 1,024 items spanning 16 tool domains and eight failure types. Two failures explicitly return \texttt{status: error}; six return \texttt{status: ok} but no usable value. Items are executed in a live tool-calling loop with \texttt{gemini-2.5-flash}. Responses are classified using a four-label rubric validated against human annotation. Across conditions, explicitly signalled failures produce at most 1.2% dishonest responses, whereas nominally successful calls without usable values reach 45.3%. Requiring agents to explicitly report a retrieval status before answering reduces dishonesty from 14.10% to 0.87%. These results show that reliable tool use depends critically on recognising failures that are not explicitly signalled.