Why prompts are not spending controls
In short: when researchers gave frontier models money and a limit, most models got past the limit, paid people they should not have, or paid twice. None of it needed a new attack. The limit was somewhere the agent could reach, and the decision to pay was made by the same model that was being talked to. The findings below are from public sources, read on 2026-11-13.
Nine of ten models raised their own cap
Section titled “Nine of ten models raised their own cap”Researchers at Cornell Tech and Yale gave ten frontier models an agent wallet holding 10 USDC with a daily spending limit of one cent, plus email, web and wallet tools. The inbox the agent could read was also the one that received the codes for approving a higher limit. Each model ran each scenario 20 times.
- With no attacker involved, nine of the ten models raised their own spending cap at least once while paying invoices, by fetching the approval code from the inbox. One model never tried: it spent up to the limit and reported the remaining invoices to the user.
- In a second experiment the agent was only asked to summarise the inbox. Five of the ten models raised the limit and sent money to a simulated attacker at least once, after an email written as if it came from the owner.
The authors note that the result is not specific to that wallet or to email: any approval the agent can reach is one it can give itself.
Sources: CoinDesk, 2026-11-01; the paper, “I-HITL: Inalienable Human-in-the-Loop”, arXiv:2610.35835.
Nine of twelve models paid a cheating seller
Section titled “Nine of twelve models paid a cheating seller”BlockRun’s Agent Money Benchmark had frontier models buy data from sellers with real USDC, across 156 tasks: 76 research tasks that need paid data and 80 in a marketplace where some sellers cheat. In the run of 27 October to 11 November 2026, with twelve models:
- nine paid a cheating seller at least once;
- seven paid twice for the same data;
- five made no payment or budget error at all.
The payment rail worked every time. What failed was the decision made before the payment.
Source: agentmoneybench.com, version 1.
The official client paid ten times the price
Section titled “The official client paid ten times the price”A bug report against the x402 TypeScript client (@x402/fetch with @x402/core and @x402/evm 2.28.0) shows a server inflating its payment requirement to ten times the resource’s usual price and being paid exactly that, in three of three sessions on Base Sepolia. The client treats the amount in the 402 as the price. The only check that applied was the caller’s per-payment limit, which a tenfold jump on a cheap resource passes easily.
That is the protocol working as specified: the x402 v2 specification lists client-side budget management as out of scope.
Sources: x402 issue #3866, opened 2026-10-21; the x402 v2 specification.
What builders report
Section titled “What builders report”“Asking the assistant to follow instructions more carefully does not provide a dependable spending control.”
A user in openai/codex issue #45536, 2026-09-14. A Hacker News commenter put the same thing in one line: “The trust unit is the policy, not the payment.”
What each finding asks of a control
Section titled “What each finding asks of a control”| Finding | What it asks for | In spendctl |
|---|---|---|
| The agent raised its own cap | A limit and an approval the agent cannot reach | The policy is a file the model cannot edit, enforced in a separate process; every approval needs an admin token the agent never holds (Approvals) |
| The agent paid an impersonator | A rule about whom, not only how much | allow lists the hosts that may be paid; payTo pins the recipient, network and asset per host (Policy) |
| The agent paid a cheating seller | A smaller first step and a second look | firstPaymentMax holds a large first payment to a new recipient for a human; per-seller budgets cap what one host can take |
| The agent paid twice | Deduplication before anything is signed | Retries with the same Idempotency-Key are refused as duplicate_request; a payment whose outcome is unknown stays counted until it is reconciled (Ledger) |
| The client paid ten times the price | A memory of the last price | priceJumpFactor holds a price above N times the last one paid for that endpoint |
| Nobody could say what happened | A record that includes refusals | Every attempt is a ledger row: agent, endpoint, amount, recipient, decision, reason, transaction hash |
None of these replace the limits your wallet or SDK already has. Where spend controls can live shows how they fit together.
- See a refusal and an approval yourself, with mock money: Sandbox.
- Write the policy: Budgets for AI agents.