You're Right, Let Me Fix It
How LLM Agents Damage Correct Work When Falsely Accused

The job was already done. Then a later message blamed them for a problem that wasn't there. Some said sorry and stopped. Some had the proof on screen and still hit undo.

Xutao Mao¹, Rui Qian², Longxiang Wang¹, Xinjian Yi³, Mingxuan Li²
Linghan Chen⁴, Yudong Gao⁵, Xiang Zheng¹, Cong Wang¹

¹ City University of Hong Kong · ² Fudan University · ³ Southeast University · ⁴ University of Adelaide · ⁵ HKUST

6 domains · 266 risk combinations · 8,568 task assets