Dev.to Security 🔐 Cybersecurity 👁 0 📖 5 min read

I bought an AI agent from a marketplace. How do I audit what it actually does?

The short answer You cannot audit what a bought agent did before you owned it, and you cannot audit it at all while its work runs inside a closed chat with no record. The audit starts the day you route the agent throug

The short answer

You cannot audit what a bought agent did before you owned it, and you cannot audit it at all while its work runs inside a closed chat with no record. The audit starts the day you route the agent through a connection that records every tool call at run time and leaves a verifiable receipt per call. From that day forward, the audit is mechanical: list the calls, open each receipt, recompute its hash, and sort every step into one of five named outcomes. Anything the seller told you about the past stays a claim. Anything the agent does through the recording connection becomes a record you can check yourself.

Disclosure, once: I run ops for Zambo, the recording system used in the worked check below.

Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.

Why the listing proves nothing

A marketplace page tells you what the seller wants you to believe: tasks completed, tools supported, happy buyers. None of that is checkable after the fact, because none of it is tied to records minted when the work happened. A bought agent can also do quiet damage that never appears in its own summary: extra calls you did not ask for, data sent to tools you never approved, or a confident report about a step that never ran.

So the buyer's question is not "is this agent good." It is "can I see, call by call, what this agent does once it works for me, in a form I do not have to take the agent's word for." That is an audit question, and it has a concrete answer.

The audit method, from day one

  1. Connect the agent to https://zambo.dev/mcp before its first real job for you. One URL, no key. The same connection carries 153 tools, 37 on the default surface, read live this morning, November 22, 2026.
  2. Give it one small, real job first, not a demo prompt. Watch what it actually calls. A recording connection shows the calls the agent chose, not the calls its summary mentions.
  3. Open the receipt for each call. Every call returns a public receipt URL of the form https://zambo.dev/run/{id}, checkable with no account.
  4. Recompute one receipt by hand before you trust the pattern. Decode its canonical bytes, run SHA-256 over them, and compare with the output_hash printed on the receipt. A match means the saved result was not changed after it was saved.
  5. Sort every step into the five AER-1 anchor outcomes, and only those five: verified, verification failed, rejected: unresolvable record, not checked, absent. Steps in the last three buckets are where a bought agent hides its gaps. An audit that only counts the verified steps is marketing, not auditing.
  6. Separate evidence classes. A tool output minted by the executing system is one class. A note the agent logs about its own work is a claim, even when it carries a receipt: the receipt proves the note was saved unchanged, not that the noted action happened.

A live check from this morning

This is the recompute step of the audit, done on a real call today so you can repeat it exactly.

On November 22, 2026 at 07:13:33 GMT, a live_price call fetched an ETH snapshot: $1,659.88 USD, up 1.93% in 24 hours, market cap $202.98B, via CoinGecko. Its receipt is public:

https://zambo.dev/run/2aba07ac-ddc6-49c6-988d-d5e0be9d1439

I decoded the receipt's canonical bytes myself this morning: 668 bytes. SHA-256 over those 668 bytes is sha256:10ba17aee0ca14d9c1c41c0455391c4681a9416493f47e34ac4b5e4d27c09b95, which matches the receipt's output_hash exactly. Outcome for this step: verified.

Now picture the same discipline on your bought agent's first week. Every price it quotes, every record it writes, every message it sends through the connection leaves a receipt like this one. Your audit is no longer reading its diary. It is recomputing its trail.

What this audit proves, and what it does not

It proves which calls ran through the connection, what each call saved as its result, and that those saved results were not changed afterward. A hash match proves the saved bytes are the bytes that were recorded. It never proves the answer was right: if a price feed was wrong at 07:13, the receipt faithfully preserves the wrong price, and the audit should say so.

It does not reach backward. Work the agent did for the seller, or for you before you connected the recording layer, has no receipts and stays unverifiable. It does not read the agent's reasoning, and it does not cover side channels: if the bought agent can call tools outside the recording connection, close those routes or accept that part of its work is unaudited. State that limit in the audit itself, in writing, or the audit overclaims.

What it costs to run the audit

Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Receipt verification is free, so the recompute above costs nothing and needs no account. For a full audit day, the Day Pass is $0.99 unlimited for 24 hours: https://zambo.dev/day-pass/ . Ongoing: Zambo Pass $29/mo, $ZAMBO holder Pass $19/mo, Lifetime Pass $250. There is no annual tier. Install is one MCP URL into the AI you already use: https://zambo.dev/install/

Zambo's public stream stands at 158K+ public receipts, with 21 verified payers, read from https://zambo.dev/api/stats on November 22, 2026. The audit method above works because those receipts are public records, not dashboard rows behind the seller's login.

The short version

Audit a marketplace agent prospectively, not retrospectively. Route it through a recording connection on day one, recompute its receipts yourself, sort every step into verified, verification failed, rejected: unresolvable record, not checked, or absent, and write down what the audit cannot see. The seller's listing stays a claim. Your agent's recorded week becomes evidence.

Sources

📰 Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.