arXivBo Lv, Zhiheng Xu, KeDong Xiu, Ruyi Ding, Tianhang Zheng, Zhibo Wang, Kui RenSat, May 23, 2026, 7:06 PM PDT
score 15.9
Detecting unsafe AI without reading user prompts using GPU patterns
Original: RouteScan: A Non-Intrusive Approach to Auditing MoE LLMs Safety via Expert Routing Telemetry
Source: arxiv.org ↗
Writing ELI5 summary…