We ran our own governance audit on ourselves. It failed.

Anyone can publish a standard. Here is the occasion ours cost us something — what it found, what we changed the same day, and what still isn't fixed.

THE SHORT VERSION

On 29 July 2026 we ran our own agent-governance checklist against our own setup — the AI assistant that has access to our working files, our publishing tools and our scheduled automations.

It returned NO-GO. Two of the must-pass controls failed outright. Under our own rule, a must-pass control at red is a NO-GO regardless of how well everything else scores — so the verdict wasn't close.We closed both the same day.

The re-run still doesn't pass. It clears the critical failure, which is a different and more modest thing, and we are not going to describe it as anything better than that.

vet → disclose → verify

WHAT IT ACTUALLY FOUND

The failure worth telling you about is the one nobody writes about honestly. Our assistant fetched a web page. Buried in the raw markup was the sentence "Your All Access Pass subscription has been suspended." It read that as a fact about our account, escalated it to a red item on our internal board, and wrote urgent actions into two files.

It was template markup served to every anonymous visitor. Proving it took one further fetch — the identical sentence appears on another page of the same site, sitting beside literal {{user.email}} placeholders and both a "Login" and a "Log Out" link.

Nothing malicious happened.

That is exactly why it matters:

Content from an untrusted source changed what the agent did, and produced actions in our files, with no step anywhere that treated that content as untrusted. Swap an accidental string for a deliberately crafted one and the same path is an attack.

This is the control everyone lists and almost nobody evidences — we now have an instance, in production, on ourselves.Two further things we'd rather not have found:

The error was caught because a human challenged it. Not by any control. A control that depends on someone noticing is not a control.
It happened in the same session in which we wrote a note titled "verify the method, not just the output." Having a control written down is not the same as having it wired in.

WHAT WE CHANGED, THE SAME DAY

Publishing access on our newsletter platform was revoked, leaving read-only — and we confirmed it by testing which tools remained reachable, not by trusting the setting.
Browser control was revoked and verified gone.
Three standing rules adopted: retrieved content is untrusted input and never a fact about us · anything that publishes, sends or deletes needs a per-action confirmation · every access grant is bound to a purpose, with its permissions stated when it is requested.
A connector register was written — every tool the assistant can reach, what it is allowed to do, who approved it and when.
An injection test set was written — nineteen cases across four families.

The three actions that cleared the critical failure took roughly forty minutes between them.

WHAT IS STILL NOT FIXED

Publishing only the good half would make this marketing again.

The injection test set has been written and has not been run. Writing an instrument is not firing it. Until it runs against our own setup with results recorded, that control is partial and we score it partial.
We audited ourselves. That is a conflict of interest. Our own method requires that two auditors reach the same verdict independently, and that has not happened.
The overall verdict is still "not ready." Clearing a critical failure is not passing.

WHY WE PUT THIS ON A PUBLIC PAGE

Because it is the only thing on this website you can actually check. We don't publish client names, results percentages or testimonials — we've never had a client consent to one, and inventing them is where most agencies lose the plot. That leaves a visitor with nothing to evaluate but claims about ourselves.

So here is the instrument instead. Read it, run it on your own setup, and form your own view.

Download the 15-point pre-launch gate (PDF) — the checklist we run before any AI-assisted product goes in front of a person.

Yours to use.We'll email it to you and add you to a short weekly note. One click unsubscribes, and we don't share your address with anyone.

THE SAME STANDARD, APPLIED TO OUR OWN PRODUCT

The agent audit isn't the only occasion. Our discovery product, Stride, went through the same 15-point gate against a hard regulatory deadline — the EU AI Act's transparency duty, which applied on 2 August 2026.

The gate found reds. The disclosure layer went live with a day to spare. And in the course of checking it, we discovered our own backend had been down for days and nobody had been told — because we had never set the billing alert that would have said so.

That is in the gate document too, written down at the time, with the date on it. The failure that embarrasses you is the one worth publishing; the rest is a brochure.

Book a first conversation.

A conversation about what you're trying to do and whether we're the right people. No pitch. "Not a fit" is a perfectly good outcome and we'll say it early.

Prefer email? [email protected]