Anthropic's move to issue employee badges and assign permissions to AI sounds nice, but Kenton Varda is right—when it really scales up, who takes the blame?

CoinWorld News reports that Anthropic has introduced an agent identity security mechanism for Claude Tag, assigning independent permission packages to each channel and directly allocating a dedicated AI account exclusively, to eliminate privilege-escalation and data-leakage backdoors in multi-person collaboration. It adopts dynamic policy configuration to replace borrowed credentials. Security expert Kenton Varda criticized the approach as difficult to handle large-scale tasks and undermining human accountability, arguing for a capability-based security model to ensure that AI actions can be traced back to specific initiators.