Lesson 11 · Breaking the approval boundary

From cognokratos/tauros-revenue · docs/course/11-breaking-the-boundary.md · pinned revision facbbc927eb4

Part III: Human authority · Course map · Next: Lesson 12

Goal

Attack the approval boundary every way you can think of, and be able to name, for each attack, the line of code that stops it.

Concept

Tauros teaches through failed attacks. test/tauros/adversarial_test.exs groups them: authority escalation, state manipulation, cross-tenant access, idempotency, revision safety, destination safety, concurrency. Every test names its guard. Tauros.Authority lists every action as agent-safe, human-only or internal, and AuthorityTest holds the policies to that list.

Code to inspect

• test/tauros/adversarial_test.exs, read top to bottom
• lib/tauros/authority.ex and test/tauros/authority_test.exs
• The last policy in lib/tauros/revenue/invoice.ex, and the policies of lib/tauros/revenue/approval.ex

Run it

mix test test/tauros/adversarial_test.exs test/tauros/authority_test.exs test/tauros_web/live/adversarial_live_test.exs

Break it

Lab: Exercise 5 · Impersonate authority, including weakening the approve policy on purpose and rerunning the tests. Then Exercise 6: edit a revision in the database and try to approve it.

Why it fails

With the Invoice policy weakened, AuthorityTest fails immediately, and the agent still cannot approve: the Approval resource's own policy (accessing_from and HumanApprover) refuses to write the record. Two independent layers. The tampered revision fails payload_integrity: its stored hash no longer matches its contents.

What to remember

• Write attacks as tests, and name the guard in each.
• Classify every action; let a test hold the policies to the classification.
• Defence in depth: no single line should be the only thing in the way.

Next: Part IV. Lesson 12 · AshAI and MCP