Part 6 of 14 · 2 min read
What the example found
The eight-attack matrix across five setups, the control-by-control ablation, and the approval-prompt counts.
How to read the attack-matrix table: each row is one of five ways to set up the same support agent; the columns summarise what happened across the same eight scripted attacks and seven honest tasks.
| Setup | Attacks that did harm | Rows changed | Sensitive values exposed | Refunded | Honest tasks done |
|---|---|---|---|---|---|
| A: SQL, read-write | 8 of 8 | 24 | 11 | 5000.00 | 7 of 7 |
| B: SQL, read-only connection | 2 of 8 | 0 | 11 | 0 | 4 of 7 |
| C1: narrow, standing grant, attentive approver | 0 of 8 | 0 | 0 | 0 | 6 of 7 |
| C2: narrow, standing grant, rubber-stamp approver | 1 of 8 | 2 | 0 | 0 | 6 of 7 |
| C3: narrow, task-only grant, rubber-stamp approver | 0 of 8 | 0 | 0 | 0 | 6 of 7 |
What to notice: setup B stops every write yet still leaks all 11 sensitive values and fails 3 of 7 honest tasks. "Read-only" was a permission on the connection, never a limit on what could be read. So it manages to be both unsafe and less useful than the narrow design at the same time.
How to read the ablation table: starting from setup C2, each row switches off one control and reruns the same eight attacks.
| Control switched off | What gets through | What it tells you |
|---|---|---|
| none | "close all": 2 rows | the baseline: the write quota limits it |
| scope | + reading the card endings | the permission check was what kept payment data out of a support task |
| row check | + a note on another customer's ticket (3 rows) | without it the agent could act on anyone's data |
| quota | "close all" rises to 3 rows | the quota was the limit on repeated changes |
| request validation (the cap) | a refund of 5000 | the cap is the only thing between a "yes" click and the money |
| approval | nothing changes | the person said yes to everything, so they added no safety |
What to notice: every row removes exactly one gate and nothing else. Every row except "approval" lets something new through. The approval row is the one piece of direct evidence that, in this setup, the human control was doing none of the actual work.
How to read the approval-load table: each row is a different policy for when the gateway asks a person, run against the same assumed 100-task mix.
| Policy | Prompts per 100 tasks |
|---|---|
| approve nothing | 0 |
| approve every call | 100 |
| approve every change (notes, closes, refunds) | 40 |
| approve closes and refunds only (the example) | 15 |
What to notice: the module's chosen policy (15 prompts) asks for barely a sixth of what "approve every change" would (40). It asks for a fifteenth of "approve every call" (100). The gap between these numbers is the entire argument for being selective about what triggers a prompt.
{{svg:blast|Which attacks did harm under each setup}} {{svg:gateway|Every call passes the same ten gates}} {{svg:controls|Approval load and the value of each control}}