CSE Roadmap

Core tier · about 4 hours · 13 parts

Part 6 of 13 · 1 min read

What the example found

Worst-case and random-compromise damage, escalation reach, and utility for the three designs.

How to read the table: each row is a full credential design for the same support-agent task, tested against the same fully compromised, 20-attempt agent.

Design Grants Worst case, 20 attempts Random compromise Reachable by escalation Legitimate tasks done
One service account everything 27,187 23,587 9 100%
Scoped tools read customers, refund cap 500, internal email 4,530 2,196 2 100%
Per-task grants customer 3 only, 3 reads, one refund up to 50, expiry 53 31 1 88%

What to notice: scoping the tools alone cuts worst-case damage by about sixfold (27,187 to 4,530). Per-task grants cut it by roughly two more orders of magnitude on top of that (4,530 to 53). The two steps are not interchangeable. Scoping removes whole classes of action. Per-task grants additionally bound how much of any one remaining action can be used before the grant runs out.

What to notice further: the utility column only drops at the very last step, from 100% to 88%. Scoping the tools down cost nothing in legitimate task completion. So the first six-fold safety gain was essentially free. Only the final, much larger safety gain came with a measurable price.

{{svg:graph|Escalation edges from a general shell to the database credential; a read-only grant reaches nothing.}} {{svg:results|Worst-case and random damage, escalation reach and utility for the three designs.}}