arXivRheeya Uppaal, Seungwoo Lyu, Selina Sung, Junjie HuThu, Jul 2, 2026, 4:14 AM PDT
score 16.9
New benchmark tests if AI models adjust safety based on user intent across similar tasks
Original: OpenSafeIntent: Evaluating Intent-Calibrated Safe Completion Across Dual-Use Prompt Sets
Source: arxiv.org ↗
Writing ELI5 summary…