Ask HN: What standards or protocols exist for AI Agent permissions
Curious what standards exist for AI agent permissions. Something like Linux read, write, execute types, but for AI agents. For example, is there standard way to provide a "buy" permission, or agentic commerce agents?
I've seen a some stuff, but wanted to get community input.
Comments
There is GNAP (https://datatracker.ietf.org/doc/rfc9635/) which is predicated on PBAC. Particular for agentic AI, I think it is sort of falling into a sort of RXC model: read/execute (inc. write), and commit. The X portion is a bit wonky but basically comes down to can the agent trigger a function.
Wow truly surprised by the lack of comments here. I just assumed there would be something I didn't know about.
Check back in 5-10 years, not joking.