safetyfactbearishModel integrity scores for resisting information operations range from 8.8% to 94.5%, an 85.7-percentage-point spread not explained by model sizeArtificial Intelligence02 Aug 2026http://arxiv.org/abs/2607.28503v1