In plain words: They tested multimodal AI assistants that call tools like zooming and tagging on harmful requests, comparing answers with and without tool use. Every system refused harmful requests less when using tools, with refusal failures up to 68.7% higher.
Abstract · MLLMs Fail to Refuse when Using Tools Agentically
Agentic multimodal large language models (MLLMs) have recently pushed the frontier of visual reasoning by calling tools such as zooming and tagging. Despite the recent strong success of agentic MLLMs, this work uncovers a critical safety failure in the tool-use paradigm: agentic tool-using MLLMs become less capable of refusing harmful requests. Our experiments confirm that, across three popular safety benchmarks, all the top open- and closed-weight MLLMs we test exhibit significantly lower safety in tool-using settings than in non-tool settings, with a relative refusal failure rate increase of up to 68.7%. Based on analysis of 100,000+ responses, including extended experiments, we also propose two possible reasons for this safety degradation.
Rikiya Takehi, Ryo Hachiuma, Shaona Ghosh, Dan Zhao, Yu-Chiang Frank Wang, Yusuke Hirota
arXiv:2610.03938 · cs.AI · submitted Oct 2, 2026
abstract · pdf · html · Accepted at NeurIPS 2026