AI compliance guard models and activation probes cannot read the rules they enforce, a new arXiv audit shows. Deleting or ...