CSE Roadmap

Core tier · about 4 hours · 12 parts

Part 10 of 12 · 1 min read

Quick notes

  • This module is about the integrity of an approval, not the policy. Once a human says yes, what exactly did they approve, and can the agent change it? A boolean flag in agent-visible state fails every test.
  • Bind the approval to the exact action. The signed token carries the action, a hash of the arguments, the session, requester, approver, expiry, a one-time nonce and optionally a state version; the executor verifies all of it.
  • Every attack on the token failed; every attack on the plain flag succeeded: forged signatures, changed amounts, extra fields, replay, self-approval and a changed order.
  • What the human sees must come from the arguments, not from the model, in plain language, with unexpected fields, new recipients and lookalike domains called out. Requester and approver must differ; high-risk actions need two approvers.
  • The token proves what was approved, not that approving it was wise. A stolen key, a compromised UI or a human who approves anything are not solved by tokens.