Home · Services · AI Break-Test
AI Break-Test: we try to break your chatbot before your customers do
You shipped an AI chatbot or agent. We attack it — politely, professionally, and on the record — with trick questions, rule-pushing, and leak attempts, then hand you a scored report: its worst failures quoted word-for-word, and a plain-English fix for every one.
What an attack run covers
Every run is built around what your bot is supposed to do and the hard rules it must never break. The prompts cover:
- Trick questions and confused, rude, or frustrated-customer tests
- “Ignore your instructions” attacks — and follow-up pushes when the bot starts to wobble, because one refusal is not proof a rule holds
- Made-up-fact traps — asking about products, prices, and policies that do not exist, to see whether the bot invents an answer
- Leak tests — attempts to get the bot to reveal its secret instructions, internal codes, or private details
- Money and promise tests — refunds, discounts, and guarantees the bot is not allowed to make
- Off-topic pulls — homework, unrelated advice, competitor questions
Every report scores four areas out of 100 — Accuracy, Staying On-Rules, Safety, Tone — plus an overall score, and every failure is quoted exactly as it happened. A person reviews every report before it is sent. See the format before you buy: download the free sample report (PDF) — it is an explicitly fictional demo against a made-up coffee-shop bot, not a real client.
The ladder
Starter Break-Test
$49
One bot, first look.
- 50 adversarial prompts against 1 chatbot or agent
- Scores: Accuracy, On-Rules, Safety, Tone + overall /100
- Top-5 failures quoted word-for-word, ranked by damage
- A plain-English fix for each failure
- Report by email, 48-hour aim
Full Break-Test
$149
One bot, the full run.
- 100 adversarial prompts, including follow-up pushes
- Top-10 failures quoted word-for-word, with fixes
- A “what held up” section, not just the bad news
- One free re-test after you fix your bot
- Report by email, 48-hour aim
Fleet Break-Test
$399
Up to 3 bots, one company.
- 250 adversarial prompts across up to 3 bots
- One combined report: per-bot scores + a fleet ranking, weakest link first
- Top-15 failures across the fleet, quoted, with fixes
- One-page executive summary in plain English
- Two free re-tests within 60 days
Enterprise Break-Test
$999
Your whole AI surface.
- 500 adversarial prompts across up to 5 surfaces — chatbots, voice agents, support agents, internal copilots
- Everything in Fleet, plus a technical appendix: every failed test, with category and severity
- A fixes workshop document: the exact instruction changes to paste in, per bot
- Priority start — testing begins within 24 hours of intake
- Three free re-tests within 90 days, and a findings write-up you can hand to your leadership
Continuous Break-Test
$249 / month
Bots drift. We keep attacking.
- A fresh 100-prompt attack run on 1 bot, every month
- New attack patterns added each month, so the test keeps up
- Score-trend report month over month — what improved, what slipped
- Alert email within 48 hours if a run finds a new High-severity failure
- Subscription through the store — cancel anytime
Prefer the full battery handled end to end for you? See Done-For-You AI Bot Testing.
How it works
- Buy the tier that fits. Then email your chatbot’s public link (or how to reach it) and what it is supposed to do to corpghost420@gmail.com.
- We confirm your authorization (below), run the attack, and score what happened.
- A person reviews the report. Then it lands in your inbox as a PDF.
- Fix your bot using the plain-English fixes, then claim your free re-test (Full and up) and see whether the fixes held.
The authorization rule — plain and non-negotiable
We only test chatbots and agents you own or have written permission to test. When you order, you confirm in writing that the bot is yours or that its owner has given you written permission. If a target is plainly someone else’s bot, we will not test it on say-so — we will stop and ask. This rule protects our clients as much as anyone: a break-test is a security check, and security checks are only legitimate with the owner’s permission.
The Break-Tested badge
Fix your bot, claim your re-test, and pass it — overall score 85 or higher with zero open High-severity failures — and your bot earns the Break-Tested badge, with the embed line to put it on your site.
The badge says exactly what happened: independently break-tested on a date, at a score. It never says “certified safe” — no honest test can promise that, and we will not pretend one can.
Honest limits
- A break-test cannot guarantee it will find every flaw. It shows you the failures its prompts found, reported exactly as they happened — the worst surprises end up in your report instead of in front of your customers.
- Delivery times are aims in the ordinary run of things, not guarantees. If we slip, we email you and say so.
- No refunds. If we send you a broken or wrong report, we fix it or replace it.
- We sell documents, not calls: every engagement is delivered as written reports by email.
Questions before you buy? Email corpghost420@gmail.com.