Authority labels cut agent memory violations to zero

AuthMem-Bench found authority collapse in 48 of 49 tested agent-memory configurations, while its end-to-end experiment cut observed unauthorized actions from 16.9% to 0.0% by predicting and preserving authority labels. The controlled benchmark does not prove the defense closes every open-ended attack, but it sharpens the case for explicit authorization at agent infrastructure boundaries: builders should store who authorized a memory and what it may control, not just the remembered claim.