Secure

Most AI agent failures do not look like failures.

They can look like normal operation, right up until a system has quietly moved beyond what anyone actually authorised. We check that your AI only does what you agreed it could, like drafting quotes but never sending them, and that nothing since has quietly changed that. Whether or not you run our code, startup or established business, the review does not change shape.

You get: a written report on what your system can actually do against what it was approved to do, problems ranked by what matters most, and a plain note on anything without an easy fix.

Most engagements start with an Agent Readiness Review: one agent or one workflow, $1,200 incl. GST, and a written verdict. Secure is quoted from what it finds.

What this actually means

Say Build shipped that quote assistant. Here is what keeping it secure means.

Once it is running, that assistant is handling real customer details: names, addresses, phone numbers, sometimes payment information to send an invoice. Secure means it can only see the customer record for the quote it is actually working on, not your entire client list. It can send a follow-up email, but it cannot go digging through payment history or quietly change what someone was quoted. And if it ever tries to do something outside that, reading data it should not, sending something to the wrong person, that is caught immediately, not discovered three months later when a customer asks why their details turned up somewhere they should not have.

That protection does not come from asking another AI to keep an eye on the first one. Most AI security today works exactly that way, a classifier or filter deciding whether an output looks safe, and it does not actually protect you: whatever can fool the AI you are running can usually fool the AI watching it too, because they are built the same way and share the same weaknesses.

Our approach is different, and it is not a slogan. Every check runs on fixed rules: a rule either matches or it does not, and it cannot be talked around or reasoned with. Governance is the seatbelt, not the destination makes the longer case.

This is not a new idea we are trying out on you. It comes from the same discipline behind everything we build: work out what could actually go wrong before it happens, not after, and have someone independent check the work rather than trust our own read of it. That discipline predates AI by decades; we are applying it here, not inventing it here.

The gap this closes

Confidence and evidence are not the same thing.

Gravitee’s State of AI Agent Security Report 2026 found organisations report 82% confidence in their ability to govern AI agents, while on average only 47.1% of deployed agents are actively monitored or secured. That gap is not a knowledge problem. It is what happens when nobody has checked which one you actually have.

We work the same way for ourselves as we do for clients: find real problems and write them down plainly, including the ones without an easy fix, instead of rounding them off into reassurance. See a real example, published in full if you want to see what that deliverable actually looks like, not just hear us describe it.

For agents already live

Governance & Drift Audit

For startups and small businesses running a long-lived agent in production, or about to.

A check on whether your AI system is still doing only what you approved it to do, not what it has quietly drifted into doing since. Not a quick prompt review, not a vague risk score.

Start a Governance & Drift Audit
Deliverables
  • Checked against the three ways systems quietly drift: the same instructions coming to mean more than they did on day one (instruction drift), approvals becoming routine as everyone gets used to the system (autonomy accumulation), and one AI passing another an instruction that claims to come from a person when it did not (authority laundering)
  • A look at the record your system keeps of what it decided and did, and whether it would actually let you check its work
  • Remediation roadmap ordered by what is actually exploitable first
  • Written report and optional structured walkthrough

Pressure Test. A quoted follow-on after a Readiness Review: we attack your system on purpose and write up every weakness we find, including the ones without an easy fix. See the one we ran on our own gateway.

What this starts from

A Readiness Review first, then a quote.

Secure work starts with the free Scope Check and an Agent Readiness Review at a set price. The review shows what your system can actually do, and the Drift Audit or a Pressure Test is quoted from what it finds, because the size of your system sets the size of the job. And if real depth exceeds what we scoped together once we’re into it, we pause and tell you before it costs you anything more, not after.

Where to start
See the Readiness Review Talk about your system
A boundary stated upfront

Visibility depends on the deployment.

If your AI runs through a provider’s online service, we check everything around the model: what it can reach, what it actually did, and the rules in between. Looking inside the model itself, at what is happening as it reasons (activation-level analysis), needs a model running on hardware you control. We will state what can and cannot be assessed before work begins.