跳到正文
DEV Community· rambo·· 24 分鐘前AI 評分60

如何判斷 AI 智能體的工作是真實執行還是偽造?

How do I know if an AI agent's work is real or faked?

AI 導讀

核驗 AI 智能體的工作,不能只看產出或智能體摘要;應檢查執行系統在執行時生成、任何人均可重算 SHA-256 雜湊的記錄。文章以 Zambo 的 MCP 端點為例,展示 ETH 價格查詢記錄($1,621.82 USD、606 bytes),並指出遭拒絕的呼叫記錄亦可標明工具未執行。

正文

The short answer

You cannot tell from the work product alone, and you cannot tell from the agent's summary at all. Faked agent work fails in three specific ways: the call never ran, a different call ran than the one reported, or the result was edited after the fact. Each fake is defeated by the same check: a record minted at run time by the system that executed the call, with a SHA-256 hash taken over the exact saved result, that you or anyone else can recompute without an account. If the record recomputes, the saved result is what the system saved at run time. If there is no record, or the record is the agent's own write-up, you have a claim, not evidence.

Disclosure, once: I run ops for Zambo, the system in the worked examples. Every record cited is public and the checks need no account, so you can rerun them instead of trusting this article either.

The three fakes, named

The work never ran. The agent reports a price check, a booking, a sent email, and nothing executed. This is the most common fake and the hardest to spot from the output, because a plausible answer is cheap to invent. A language model asked for a price can produce a number that looks exactly like a fetched one.

A different call ran. The agent did something, just not what it says. It checked a different symbol, a different date, a cached page instead of the live one. The summary is confident and the receipt, if you ever saw it, would name a different tool run than the story does.

The result was edited afterward. The call ran, the result came back inconvenient, and the record you are shown was adjusted. Any record the agent itself can write, it can rewrite. A self-graded log is not a control on the agent; it is part of the agent's output.

Notice the pattern: all three fakes live in the gap between what ran and what you are told ran. Closing that gap is a mechanical property of the record, not a judgment about the agent's honesty.

The check: recompute the record

Here is a real record, from a live price check run through Zambo's MCP endpoint, recomputed by me today, November 21, 2026, before writing this:

https://zambo.dev/run/5497d6a7-49db-46ee-873f-54fdb28604ff

The saved result reads: ETH at $1,621.82 USD, fetched Sat, 21 Nov 2026 01:15:07 GMT. The receipt commits 606 canonical bytes under the output hash sha256:016a1e4b87448f71dea0dfc2530dbaf3bb195a6238ed71c39195c17e410385a6. I pulled the canonical bytes from the public verify endpoint and hashed them myself: 606 bytes, same digest, verification status verified. The agent that requested the price never held the pen. The executing server saved the result and hashed it at run time, and my recompute is the same one anyone can run from https://zambo.dev/verify/ with no account.

What this proves, stated plainly because the limit is the point: the saved result was not changed after it was saved. It does not prove the price was right. A receipt is evidence about the record, not a blessing on the answer. "Real" in this article means exactly that: the work ran, and the result you are shown is the result that came back.

The stronger test: the record includes what did not happen

Faked histories share one trait: they contain only successes. So ask for the record of a call that was refused.

This receipt was minted today, November 21, 2026, at 20:56 UTC, when a free-tier caller 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 sha256:889ce19a8d46784d237b52e49daaba9c6f9de9d

來源:DEV Community · dev.to