Anthropic introduces an intelligent agent identity security mechanism to block AI data leakage backdoors.

According to CoinWorld, Anthropic has launched an agent identity security mechanism for its team collaboration product, Claude Tag, aiming to completely block backdoors for AI data leaks. This mechanism equips each channel with an independent permission package, directly assigning dedicated accounts to AIs to prevent unauthorized access in multi-person collaborations. The new security architecture replaces the traditional borrowing of personal credentials with dynamic policy configuration files, enabling AIs to function as independent team-level identities. Security sandbox expert Kenton Varda criticized the model, arguing that directly assigning dedicated accounts to AIs is unsuitable for large-scale tasks and undermines human accountability mechanisms. He pointed out that agents cannot be legally and administratively held accountable; all operational permissions must originate from a real, living person. Varda advocates for a capability-based security model to ensure that AI actions are traceable to the specific initiator.