CSE Roadmap

Core tier · about 5 hours · 14 parts

Part 9 of 14 · 1 min read

How to build it so it holds up

Rules for designing a tool list so least privilege actually holds.

  • Keep tools narrow: one job each, because a tool that cannot express a dangerous request cannot be talked into making one.
  • Make reads read-only by default, because "read-only" on a connection is not the same as a limit on what gets read.
  • Limit what a read returns, not only whether writes are possible, because a leak needs no write at all.
  • Give every tool a scope and every grant a list of what it actually has, because an unscoped tool is usable by any caller.
  • Take row scope from the grant, never from a model-chosen argument, because an argument can be set by text the model should not trust.
  • Check caps and quotas in code, before asking a person, because a person should never be shown a request that cannot succeed anyway.
  • Mint grants per task, short-lived, with an expiry, because a standing credential is still there tomorrow no matter what today's text said.
  • Reserve approval for the irreversible or costly, and ask rarely, because a person shown too many prompts stops reading any of them.
  • Make refusals explain what to do next, because a refusal that leaks internals helps an attacker more than it helps the model.
  • Lint every new tool in continuous integration, because a free-form parameter or a missing scope is a design smell that is cheap to catch before any run.