讓 AI 智能體替我花錢安全嗎?關鍵在三項系統機制
Is it safe to let an AI agent spend my money?
AI 智能體代為支出是否安全,取決於系統是否設有智能體不能自行提高的支出上限、可由第三方重算的每筆交易紀錄,以及付款前留存的批准紀錄。Zambo 免費層的共享每日上限曾在工具執行前攔截請求,作者核對公開收據後確認雜湊相符;另一筆 x402 酒店預訂報價的雜湊也經重算吻合,但結算規格有一項 -17 數值欠缺推導公式。第三項做法是把客戶或負責人的決定記錄在交接頁上,並在款項轉移前完成批准。
The short answer
Letting an agent spend is not a trust decision. It is a mechanism decision. It is reasonably safe when three things hold at the system level, not in the agent's instructions: the spend runs into a cap the agent cannot raise on its own, every spend leaves a record a third party can recompute, and any approval the spend needs is recorded before the money moves, not reconstructed after. Remove any one of the three and you are not running an agent with a budget. You are running an agent with your wallet and a hope.
Disclosure, once: I run ops for Zambo, one of the systems below. The records cited are public, and the checks described need no account.
Condition one: a cap the agent cannot raise
A spending limit written in a prompt is a request. A spending limit enforced by the executing layer is a wall. The difference shows up the first time an agent loops, retries, or misunderstands an instruction at 3 AM.
Here is what a wall firing looks like as a record. This receipt was minted today, November 21, 2026, at 20:56 UTC, when a caller on Zambo's free tier hit the shared daily limit and the requested tool was blocked before execution:
https://zambo.dev/run/38a410ab-597e-45e0-9549-0b3bf2916b39
Its saved result states that the tool was not executed, step status payment_required. It commits 760 canonical bytes under sha256:889ce19a8d46784d237b52e49daaba9c6f9de9d58d741bc76fd402771cec6847, which I recomputed today from the public verify endpoint: match. Two things matter here. The block happened before execution, so no spend occurred. And the block itself was receipted, so the limit is not something the vendor claims exists; it is something you can watch fire, in the same record format as a success. A cap you cannot observe is a cap you are taking on faith.
Condition two: a recomputable record for every spend
When money moves, the terms of the spend should be committed before settlement, in a form a stranger can recompute. An independent builder on the x402 protocol, hwcho310, published exactly that shape today for a hotel booking quote: the terms (buyer agent_buyer_zambo, supplier agent_supplier_hotel, price 220,000 KRW, holdback ratio 0.1, refund ratio on cancel 1.0, free cancellation until 2026-11-25T18:00:00+09:00) serialized under a published canonicalization, json.dumps with sorted keys and no whitespace, and hashed. He published the canonical string and the claimed digest, sha256:a8b66da87fb7e536d6d253fe28459fdb2c1910e52cd56c50b757b3343ecf4d67. I recomputed it today from his published string: 217 bytes, same digest, exact match. That is the standard to hold any agent-spend system to: do not tell me the terms were fair, show me a commitment I can rehash.
Honesty requires the other half: his settlement spec still has one seam, named on the thread today. His credit_delta table publishes a value of -17 for one compensated case without the formula that derives it from the uncovered amount. A value you cannot derive is a value you cannot recompute, and the fix is his to publish. Spend records in this space are getting checkable fast, and the gaps are now specific enough to name, which is what progress looks like.
Condition three: approval recorded before the money moves
The third condition is about who said yes, and when. An approval that lives in a chat thread approves nothing identifiable. The shape that works: the job is wrapped as a handover, one page carrying the goal, the steps, the receipts, and a decision the client or owner records on that page. Here is a real one, opened today:
https://zambo.dev/handover/d468118b-afc1-43e5-a70c-5e198e876a04
Its Decisions section reads "No
來源:DEV Community · dev.to