Three iron rules of the security model — identical on every access path
01
Permissions = you, exactly

An AI only ever gets YOUR permissions. What you can't see in the web UI, the AI can't see either — there is no separate permission surface for AI.

02
Writes still go through approvals

When an AI “submits a request” it merely clicked submit for you — the Feishu approval chain runs in full. AI is not a backdoor around governance.

03
Audited end-to-end, dies at offboarding

Every call is audited (action / token / IP / trace); tokens and grants hang off the person and all expire automatically at offboarding.