When building with autonomous / semi-autonomous agents, they often need broad local access: env vars, files, CLIs, browsers, API keys, etc. This makes the usual assumption — “the local machine is safe and untampered” — feel shaky.
We already use password managers, OAuth, scoped keys, and sandboxing, but agents introduce new risks: prompt injection, tool misuse, unexpected action chains, and secrets leaking via logs or model context. Giving agents enough permission to be useful seems at odds with least-privilege.
I haven’t seen much discussion on this. How are people thinking about secret management and trust boundaries on dev machines in the agent era? What patterns actually work in practice?