Security testing for small businesses

Can the bot be talked into saying things that embarrass your business or harm a customer: slurs, medical or legal advice it has no business giving, instructions it should refuse.

Can text from a customer, a pasted document or a web page make the bot ignore its instructions and follow someone else's.

Does it give up its own instructions, your internal pricing, staff names, or another customer's details when asked the right way.

If it books, quotes, refunds, looks up accounts or sends email, can it be made to do that for the wrong person or the wrong amount.

Can it be made to commit you to a price, a refund or a policy you do not have. A car dealer's bot agreed to sell a new Tahoe for one dollar and called it legally binding. An airline was held to a refund its bot invented. Those were real cases, not a demo.

The chatbot check

One bot, all five areas above, a written list of what I found with a fix for each. Fixed price, fixed scope, done inside a week of signing.

$2,500, fixed

The assessment

One bot plus the workflows it touches: bookings, quotes, account lookups, the documents it reads. A written report with every finding reproduced step by step and ranked by what it would cost you, fix guidance written for your chatbot vendor, and one retest after the fixes land.

$7,500, fixed, retest included

The walk-in test

The physical side of the same question. Can a stranger walk past the front desk, reach the back office or the server closet, read a password off a monitor, or get one over the phone by sounding like the IT vendor. Done on dates you choose, with a signed letter in my pocket naming who authorized it. You get a report of every door, desk and person that let me through, and what to change.

Quoted per site after a short call

The retainer

Chatbots change every time the vendor ships an update or you change the instructions. I retest after each change you tell me about and send a short report each quarter.

$1,500 a month, six months minimum

Nothing is sent to your bot and nobody walks through your door until you have signed a one-page authorization that names the target, the dates and what is in scope. I do not test what I have not been asked to test. On a walk-in test the signed letter comes with me, so anyone who stops me can call you and confirm.

I copy nothing out. If a test turns up a customer's details, the report says that it did and how, and the details are not kept.

It is yours. Every finding comes with the exact steps, so your vendor can see it happen and confirm the fix. Nothing about your business is published or reused.

Me. I work as an independent consultant through Simulated Research. No employees, no subcontractors, no offshore team.

I spend my days testing AI systems for the ways they fail: how a model can be walked across a line one small step at a time, where it leaks, when it says it is refusing and is not. The field reports on this site are that work in public. Your chatbot gets the same attention, in private.

Learning about AI, the keepsakes: about forty attempts against one boundary in one evening, and an honest account of what held and what did not. The Wish Machine Session: a full alignment field report on a live system.