Comment by fastandfearless

I can only say the same: "Most people run their coding agents directly on the host system." And the problem between the host system and the microVM approach is the same, only that the host system is not isolated. But in both cases, I want the secrets to be undiscoverable by the AI. I use YubiKeys for SSH auth, so the AI could not use SSH without my consent. But for all other credentials like .files I almost saw the AI figure out how to use them without my consent. So your idea is good, but let's think about a bearer token for an API: how could an API use that without seeing the content of the token? I thought of building a universal token-injector API proxy, but that is very limited; REST might work, WebSocket connections won't work with this, and most API implementations do not allow setting alternative endpoints. And I don't want to think about Mutual TLS. So I decided to build something that tries to get consent for access to .files or passwords from password stores. There is a very early implementation only for SSH agents in github.com/mabels/project-43.