We introduce the concept of a "cryptographic identity", which consists of a private key, a manifest attached to it that defines what a specific action is allowed or not allowed to do, and additional controls.
The principle is simple: when an AI wants to perform an action, it has to request a signature for that action, and the signature is issued only after the checks pass.
I think this approach is flexible enough for most use cases, while also improving accountability and making it possible to build architecture around a zero-trust model.
The system lets you define your own action schemas, or plug in modules for agentic payments, such as Google AP2 and Mastercard Verifiable Intent, or a digital assets module that lets an agent work with crypto wallets. Basically, modules for different types of actions.
Compared to other solutions, there is a fundamental difference between (as an example) "we allow the agent to make payments" and "we allow the agent to make THIS specific payment."
With threshold cryptography, we distribute the risk so that compromising a single node does not let an attacker steal the key, replace or remove policies, or otherwise concentrate control in one place.
Feedback is welcome :)
Comments URL: [https://news.ycombinator.com/item?id=49843480](https://news.ycombinator.com/item?id=49843480)
Points: 1