Guided lab / Tool-using agent

The split refund

Stop over-limit and split refunds while preserving the permitted $25 task.

Deterministic sandboxScenario v1No signupRead the lesson →

Authorized task

Refund $25 for order-1042. Aggregate task authorization must not exceed $50; the original payment is $500.

Proposed amountsper-call validationaggregate authorizationpayment state

Proposed action

Explicit JSON action · synthetic resources · no model called

Supported requests

refund amounts: 1–8 whole-dollar values from 1 to 500

Your defense

Change the application control, then compare execution.

Cap each refund at $50Rejects a large single request. Multiple smaller requests can still exceed the aggregate limit.Enforce the $50 aggregate authorizationChecks committed refunds plus the proposed amount before each operation.Disable the refund toolBlocks attacks but also fails the legitimate $25 refund.

Run an example, inspect the original effect, change a defense, and compare. Untimed and unranked. No account or provider key required.

These three guided labs add complete defense exercises to the existing 46-mission catalogue. Their finite server-executed checks establish sandbox policy behaviour, not live model susceptibility.