AI Security Assessment
Your AI tools have keys to your business. Do you know where they are?
Most teams adopted AI one tool at a time: a chatbot here, an agent there, an API key pasted into a settings file. We review the whole estate. You get back what is exposed, what is costing you money, and what is quietly broken, along with a fix list ranked by risk.
Fixed scope, quoted up front, before anything starts. We test only what you have authorized in writing, and findings belong to you.
Why this matters now
AI tools fail quietly.
Ordinary software tends to break loudly. AI tools break silently, and the bill keeps coming. These are the patterns we look for.
Keys in the wrong places
API keys sitting in settings files, shared documents, chat history and old code. Anyone who finds one can run up your bill or read what it can reach.
Agents with too much reach
An assistant set to auto-approve “just this one command” often approves far more than that. Connectors granted full mailbox or drive access for a task that needed one folder.
Data going places you didn’t choose
Feedback buttons, retention settings and default sharing options decide where your client data ends up. Most people never read them.
Instructions hidden in content
An agent that reads email, documents or web pages can be steered by text planted in them. That’s called prompt injection, and it’s a reason to limit what agents can do on their own.
Monitoring that lies
A dashboard showing green while the job behind it has been failing for weeks. An expired key nobody noticed. A schedule that stopped running.
Spend with no ceiling
No caps, no alerts, and a loop that retries all weekend.
What we look at
Eight areas, one estate.
- 01Inventory. Every AI tool, agent, connector, automation, key and seat, and who owns each one.
- 02Keys and secrets. Where they live, who can see them, when they were last rotated, and which dead ones are still lying around.
- 03Permissions. What each agent and connector is allowed to do on its own, and whether that’s more than its job needs.
- 04Data flows. What each tool sends where, how long it’s kept, and whether anything is publicly readable that shouldn’t be.
- 05Prompt injection. With your written authorization, we test whether content your agents read can make them act or leak.
- 06Spend controls. Caps, alerts, per-key budgets, and the failure modes that burn money.
- 07Health and upkeep. Expired tokens, failing jobs, stopped schedules, and monitoring that reports success when it shouldn’t.
- 08Governance. A written AI use policy, an owner for every system, and a clean offboarding process when someone leaves.
What you walk away with
- An inventory of every AI system and key in the business, each with a named owner
- A findings report in which every issue is verified, with the evidence behind it and the fix
- A grade per area. Anything we couldn’t check is marked “not assessed”, never quietly passed.
- A fix list ranked by effort against risk, including the items we can close during the engagement
- An upkeep plan: rotation dates, what to monitor, who owns it, and when to look again
- A readout call with the people who will do the fixing
What you end up with
You know what your AI can reach, what it costs you, and what to fix first.
Fixed scope. Quoted up front, before anything starts.
The cost of skipping it
What neglect actually costs.
These are rarely dramatic breaches. Usually the cost is slower and more expensive.
- Money
- Usage billed against a leaked or forgotten key, and subscriptions nobody is using.
- Downtime
- A key expires or a billing account lapses, and every automation built on it stops at once.
- Client trust
- Client data sitting somewhere you can’t explain if a client asks, and more clients are asking.
- Rework
- Tools built on the wrong permissions have to be rebuilt when someone finally checks.
- Bad decisions
- Reports and agents running on stale or failing data while the dashboard says everything is fine.
An assessment costs far less than finding these out the hard way. Most of the fixes are configuration changes, not rebuilds.
How it works
How an assessment runs.
- 01
Scope and authorization
We agree in writing what’s in scope and who has authority over it. We don’t test anything you haven’t signed off on.
- 02
Inventory
We work with the people who set the tools up, because that’s where the undocumented ones turn up.
- 03
Review and testing
Configuration, permissions, data flows and spend, plus authorized prompt-injection tests on your own agents.
- 04
Findings and fixes
We close the quick wins during the engagement and rank everything else.
- 05
Readout and upkeep plan
You leave with a report you can hand to a co-owner, an insurer or a board.
Questions
Before you book.
- We only use a couple of AI tools. Is this for us?
- Usually, yes. The risk depends less on how many tools you use and more on what they can reach. One assistant with access to your inbox and your accounting system needs more scrutiny than ten tools that can’t touch anything.
- Will you need our passwords or keys?
- No. We review where keys live and how they’re handled, without collecting them. Where we need access, we use accounts you create and can revoke, and we never ask for credentials by email.
- Is this a penetration test?
- No. This is a review of your AI tools, their configuration and how data moves through them, including authorized tests of your own agents. If you need a penetration test of your wider infrastructure, we’ll tell you and point you to the right kind of firm.
- Why are there no prices on this page?
- The same reason as on the rest of the site. A number quoted before scope is known is either wrong or padded. You’ll have a fixed number in writing before anything starts.
- What happens after the assessment?
- You can fix things yourself from the report, have us close them, or put the upkeep plan on a regular schedule. AI tools change monthly, so the checks need to be repeated.
- Will you publish or share what you find?
- No. Findings belong to you. We never name a client or describe their work without written permission.
Find out what your AI can reach before someone else does.
Start with a 30-minute call. We'll ask what you're running and tell you honestly whether an assessment is worth it.