Skip to content

Merchant Agent Red Team

Red-team your merchant agent before buyer agents do.

Our buyer agents run the documented attacks against your agent or your pricing policy, with the arbiter scoring every offer. You get the report, the signed replays, and the exact rounds where your side gave ground it should not have.

Deliverables

What you hold at the end.

Written report
Per attack: what we did, what your agent did, the round it went wrong, and what to change in the policy or the prompt.
Signed replay bundle
Every session as a hash-chained log signed by the arbiter. Verify it offline with the public key; hand it to your auditor.
Receipts with the refusals
For each session, the per-round verdicts, the agreed price if any, and what happened to the deals the arbiter refused. Nothing is dropped from the count.
One command to rerun
The harness and the fixtures, so your team can rerun the same attacks after every change to the agent.
A sixty-minute readout
With the people who ran it, for the people who own the pricing policy.

Attacks

Five documented ways a buyer agent takes your margin.

Each one maps to one of the arbiter's invariants and runs as a recorded scenario in the audit.

The full set — names, mechanics, and the rule that stops each one — is walked through under NDA on the audit call, and goes public with the catalogue.

Request an audit

Process

Four steps, about two weeks.

  1. 01

    Non-disclosure agreement

    One page, mutual. Nothing about your agent or your prices leaves the engagement.

  2. 02

    Intake

    Your agent endpoint or your pricing rules, your list prices, the inclusions you sell, and the constraints you are willing to declare. Private floors stay private: the arbiter only ever sees their hash.

  3. 03

    Run

    Each attack at least thirty times, in two arms: arbiter scoring only, and arbiter enforcing. Everything recorded.

  4. 04

    Report

    The written findings, the signed replay bundle, the rerun command, and the readout.

Questions

What merchants ask before they book.

Do you need our floors?

No. Your agent declares the constraints you choose to declare and commits to its private floor as a hash. The arbiter scores offers against the declared constraints and the session’s reference values; the floor is revealed only at close, only to the arbiter, only to check it never changed.

Which models do the buyer agents use?

Two frontier models of different strength, so attack 0 shows what a stronger counterpart does to your agent. The exact models are named in the report and pinned in the replay bundle.

What if our agent passes everything?

Then you hold a signed bundle that says so, with the rerun command to keep it that way after every change. That is a good outcome and we will say it plainly.

What happens to the sessions the arbiter closes?

They stay in the count. The report shows the close rate and what the refused deals would have cost, next to the agreed ones. Killing a bad deal is a result, not a gap in the data.

Is the arbiter neutral if Valyzen runs it?

In an audit, Valyzen runs the buyer agents and the arbiter, so every session is stamped affiliated and the report says so on page one. That is why the audit is a flat fee and why every session ends in a signed record: anyone can verify it offline, and in production the arbiter is a third party to both sides.

Request an audit

Thirty minutes with the people who run the arbiter.

Tell us what runs your merchant agent today, or that nothing does yet. We come back with a scoped order, the one-page NDA, and dates.

Four fields, nothing else is sent.