Anthropic plugs AI data leaks with per-channel permission packages
Anthropic is rolling out a new identity-based security mechanism for its team collaboration tool Claude Tag. Instead of borrowing a human user's credentials, each AI now gets its own dedicated account with dynamic permissions that change depending on the channel it's in. This prevents the kind of privilege escalation that can happen when a malicious actor tricks an AI into revealing sensitive files or system keys.
Under the old model, AIs often used shared employee accounts. An outsider could ask probing questions to access private documents or coax the AI into spilling secrets stored in local config files. In the new system, the AI automatically mounts different permission packages in different chat channels. For example, it might be allowed to write to a database only in a specific dev channel, while in a public channel it can only read. Memory is also physically isolated per channel — financial secrets never leak into a dev channel conversation. All network requests get dynamic credentials injected by a gateway; the AI itself never knows the keys, eliminating the risk of password leaks.
Future updates will add just-in-time authorization with human pop-up confirmations for sensitive actions, and will layer the requester's own permissions on top. The system will only let an operation through if both the human and the AI have the right to do it — a double-check that cuts off privilege escalation at the root.