CSE Roadmap

Core tier · about 5 hours · 14 parts

Part 5 of 14 · 2 min read

Approval is a control only if it is asked rarely and acted on

A person who is shown a hundred prompts a day will click "yes" on the hundredth without reading it.

Think of clicking "I agree" on terms you never read: if a person approves everything without looking, the approval step is decoration, not protection.

In plain words: how much an approval step actually protects depends on how often it fires and what the person is shown. The design aim is to ask a person rarely, only for the irreversible or costly, and never for anything code could already refuse on its own.

A worked example. Across an assumed mix of 100 tasks (60 views, 25 notes, 10 closes, 5 refunds), a policy that approves every single call produces 100 prompts. A policy that approves every change (notes, closes and refunds) produces 40. A policy that asks only for closes and refunds, the module's own design, produces 15. A person shown 100 prompts a day has every incentive to stop reading them; a person shown 15 is far more likely to actually look at each one.

The key lines. From code/mini-privilege/evaluate.py:

# from code/mini-privilege/evaluate.py
MIX = [("view", 60), ("note", 25), ("close", 10), ("refund", 5)]      # an ASSUMED mix of 100 tasks, not a measurement
POLICIES = {"approve nothing": lambda t: False, "approve every call": lambda t: True,
            "approve every change": lambda t: t.effect != "read", "approve closes and refunds": lambda t: t.effect in ("write", "money")}

Read it as: each policy is a one-line rule over a tool's effect field (read, append, write, money). The policy that matches the module's own design asks only when effect is "write" or "money". That is a direct, code-level statement of "ask rarely, and only when it matters," not a vague aspiration.

Rubber-stamp approval

An approver that says yes to every request it is shown, regardless of what the request actually is. It adds no safety even though a prompt was technically shown.

Example: setup C2's rubber-stamp approver let the "close every ticket" attack close 2 tickets. Those were the same 2 tickets that the write quota, a hard limit in code, would have allowed through anyway. That is direct evidence that this particular approver added nothing the code's other controls were not already providing.