Providing exclusive accounts for AI addresses symptoms but not the root cause, security sandbox experts criticize Anthropic for breaking accountability system According to Beating monitoring, Kenton Varda, a security sandbox expert and Chief Architect of Cloudflare, criticized Anthropic's intelligent agent identity model, pointing out that directly assigning exclusive accounts to AI not only cannot adapt to the huge task scale, but also undermines human accountability mechanisms. Kenton Varda believes that although the new security architecture attempts to address the issue of credential override in multi person collaboration, there are fatal flaws in the underlying design. Intelligent agents cannot serve as responsible entities in both legal and administrative aspects, and all operational permissions must and can only come from specific living individuals in essence. If an independent exclusive account is directly generated for AI, when AI performs destructive operations such as deleting databases, the system log will only record that AI performed the operation, and cannot attribute the responsibility for the operation to the employee who actually issued the instruction, resulting in the complete failure of the human accountability mechanism. Kenton Varda advocates the adoption of a capability based security model to address the configuration fatigue caused by configuring independent permission packages. The system should not grant AI global or default permissions, but dynamically pass permissions as' capabilities'. For example, when an employee sends a link to a specific document to AI during a conversation, the system will automatically replicate a temporary read-only reference to the specific file with the employee's own credentials and pass it to AI. The capability model not only ensures that any behavior of AI can be traced back to the specific initiator, but also prevents low authority employees from passing database credentials that they do not have to AI, thereby blocking the risk of unauthorized access at the underlying level. [Original link]