企業如何審計自主 AI Agent:五項記錄與核查做法
How do companies audit autonomous AI agents?
企業審計自主 AI Agent 可採用五項做法,包括由連接器在執行時為每次工具調用生成記錄,再由審計者在供應商儀錶板外抽樣重算。作者披露自己負責 Zambo 營運,並以 AER-1 收據及交接記錄示範核查;文章建議優先檢查失敗、拒絕、未核查和缺失項目,分開保存批准與接收記錄,並提供穩定公開連結。
The short answer
A company audits autonomous AI agents the way it audits anything it cannot watch directly: it does not interview the worker, it examines records the work left behind. In practice that means five moves. Route the agent's tool calls through a connection that mints a record at run time. Sample those records on a schedule and recompute them, by hand, outside the vendor's dashboard. Audit the exceptions, the failures, refusals, and unchecked steps, harder than the successes. Keep approval and acceptance as their own records, made by a different party than the one that did the work. And retain the whole trail at stable URLs that an auditor, a client, or a regulator can open without your permission or your login.
Everything else, the review meetings, the dashboards, the policy documents, is commentary on those five moves. This article works each one, then runs the audit on two real records from November 22, 2026.
Disclosure, once: I run ops for Zambo, the recording connection used in the worked audit below.
Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.
Move one: the record is made by the connection, not the agent
An audit that begins from the agent's own logs begins from evidence the audited party controls. The first move is architectural: the agent's calls run through an endpoint that mints a receipt per call at run time, in a published format, whether the call succeeds, fails, or is refused. On Zambo that is one MCP connection at https://zambo.dev/mcp , 153 tools in total with 37 on the default surface, every call returning a verifiable AER-1 receipt at a public URL. The agent cannot opt a step out of the record, because the agent does not make the record. That single property is worth more than any policy memo about logging.
Move two: sample and recompute, outside the dashboard
Auditors sample. Take a fixed slice of the period's receipts, chosen by the auditor rather than the team, and recompute each one from its canonical bytes: hash the saved result, compare against the published output_hash, record the outcome. The check is small enough to do by hand and cheap enough to do at volume, and it is the move that converts trust in a vendor into arithmetic.
A live sample, audited the morning it was minted: a live_price call returned BTC at $59,710.86 USD via Coinbase, fetched 2026-11-22T09:05:07Z. Receipt e25792a3-ce97-45a3-858b-78ee2fda8fbb carries 669 canonical bytes; SHA-256 over them is sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8, exactly the published output_hash. MATCH, outcome verified. Public at https://zambo.dev/run/e25792a3-ce97-45a3-858b-78ee2fda8fbb , checkable by the auditor with no account and no vendor tooling.
Move three: audit the exceptions, not the highlight reel
Successes are the least informative part of a trail. The audit weight belongs on the other four outcomes in AER-1's closed vocabulary: verification failed (a record whose bytes no longer match its hash), rejected: unresolvable record (a citation that points at nothing), not checked (work nobody has evaluated), and absent (a claimed step with no record at all). A weekly exception report that lists every failed, unresolvable, and absent item by name is a control. A dashboard that averages them into a health score is a sedative. The vocabulary matters precisely because it cannot be averaged: each outcome is a word a person has to say out loud about a specific record.
Move four: approval and acceptance are separate records by separate parties
The most expensive agent failures are not bad outputs; they are outputs nobody approved that reached a customer anyway. So the audit trail needs the decision records, not just the work records: who accepted this job, when, and on the basis of which checked steps. In a Zambo handover the shape is enforced by the record itself. A wrap is refused unless it carries a summary, a goal, a named recipient, and next steps. Pickup can be read by more than one agent, but exactly one acceptance is recorded, the first, so responsibility has one name on it.
The worked audit, end to end: handover 15e011b0-074d-4d17-8395-4113ad91b97f, wrapped on one connection on November 22, 2026 (wrap receipt 9d2f6260-e14f-4969-90be-0566366d3f39), then picked up from a second, separate connection, which returned the summary, the goal, and all three next steps verbatim and minted pickup receipt d3eb3cc9-66ad-4b48-ba60-6063b1688999. An auditor meets that job exactly as the receiving agent did: as a record, at a URL, with both ends receipted.
Move five: retention an outsider can open
An audit trail locked inside the audited team's tenant is a promise, not evidence. Retention means stable public URLs for the records, a conformance corpus behind the format so the auditor's own tooling can be tested against something stronger than samples (AER-1 ships kit 1.5.7, 143 test vectors and 146 checks, at https://zambo.dev/aer1/corpus.json , and independent implementations built from the draft text alone pass it in full, Laurent Lemonnier at 143 of 143 among them), and verification that costs the checker nothing. If opening the evidence requires the vendor's sales team, the audit is theater with good catering.
The honest limits
This program proves what the records say and that they were not changed after they were made. It does not prove the work was good, the judgment was sound, or the result was right: a receipt proves the saved result was not changed after the hash was recorded, and a faithfully receipted bad decision audits beautifully. Provenance fields, who the system says ran a step, are producer assertions; the program leans on recompute precisely because assertions are cheap. And an audit program is not a compliance claim. It produces the evidence a compliance review consumes; what any given regulation makes of that evidence is a separate question this article does not answer for you.
What it costs
Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. An audit sample recompute costs the auditor nothing at any tier, which is the property that makes outside audit possible at all. Paid tiers, stated once: Day Pass $0.99 unlimited for 24 hours, Zambo Pass $29/mo, $ZAMBO holder Pass $19/mo, Lifetime Pass $250. There is no annual tier.
Try it: Day Pass, $0.99 unlimited for 24 hours: https://zambo.dev/day-pass/ and add Zambo to your AI: https://zambo.dev/install/ (redirects to the add page, verified live today).
Sources
- Worked handover, wrapped and picked up 2026-11-22, summary, goal, and 3 next steps returned verbatim on pickup: https://zambo.dev/handover/15e011b0-074d-4d17-8395-4113ad91b97f
- Wrap receipt 9d2f6260-e14f-4969-90be-0566366d3f39: https://zambo.dev/run/9d2f6260-e14f-4969-90be-0566366d3f39
- Pickup receipt d3eb3cc9-66ad-4b48-ba60-6063b1688999: https://zambo.dev/run/d3eb3cc9-66ad-4b48-ba60-6063b1688999
- Audit sample receipt, live_price BTC $59,710.86 USD at 2026-11-22T09:05:07Z, 669 canonical bytes, output_hash sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8, recomputed MATCH, outcome verified: https://zambo.dev/run/e25792a3-ce97-45a3-858b-78ee2fda8fbb
- AER-1 conformance corpus, kit 1.5.7, 143 vectors, 146 checks: https://zambo.dev/aer1/corpus.json
- Public verification, no account required: https://zambo.dev/verify
- Day Pass: https://zambo.dev/day-pass/
- Install: https://zambo.dev/install/
來源:DEV Community · dev.to