Kihan launches a separate permission layer for AI agents
Kihan's proposed system checks whether an agent is authorised before a connected tool acts.
An agent's ability to do something is different from its permission to do it.
What happened
Kihan announced its formation on 5 November, introducing an architecture it calls Constitutional Computing. The design evaluates consequential agent actions against signed rules outside the model. A connected tool then checks a permit bound to the specific action before executing it.
Why it matters
This moves permission checks into a separate enforcement mechanism. A written instruction to an agent can guide behaviour, but an external check can determine whether a tool is allowed to carry out the request.
The bigger picture
The company says its decision engine works, while production enforcement integrations remain in development and initial enterprise evaluations are targeted for December. The announcement describes an approach and product plan, rather than independent proof that it prevents every unauthorised action. Its usefulness will depend on which execution paths are actually covered.