الملخص
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.
الكلمات المفتاحية
الموضوع
بيانات النشر
- المجلة
- غير متاح
- وصول مفتوح
- وصول مفتوح أخضر
اقتبس هذه المقالة
APA 7
Takehi, R., Hachiuma, R., Ghosh, S., Zhao, D., Wang, Y. C. F., & Hirota, Y. (2026). MLLMs Fail to Refuse when Using Tools Agentically. https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically
MLA 9
Takehi, Rikiya, et al. "MLLMs Fail to Refuse when Using Tools Agentically." https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically.
شيكاغو (المؤلف–التاريخ)
Takehi, Rikiya, Ryo Hachiuma, Shaona Ghosh, Dan Zhao, Yu-Chiang Frank Wang, and Yusuke Hirota. 2026. "MLLMs Fail to Refuse when Using Tools Agentically." https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically.
هارفارد
Takehi, R., Hachiuma, R., Ghosh, S., Zhao, D., Wang, Y. C. F. and Hirota, Y. (2026) 'MLLMs Fail to Refuse when Using Tools Agentically', Available at: https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically.
فانكوفر
Takehi R, Hachiuma R, Ghosh S, Zhao D, Wang YCF, Hirota Y. MLLMs Fail to Refuse when Using Tools Agentically. https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically
IEEE
R. Takehi, R. Hachiuma, S. Ghosh, D. Zhao, Y. C. F. Wang, and Y. Hirota, "MLLMs Fail to Refuse when Using Tools Agentically," https://omanscience.com/ar/articles/mllms-fail-to-refuse-when-using-tools-agentically.