AI Safety Audit

When AI goes wrong in a business, it rarely crashes. It produces a confident, well-written, completely normal-looking response that happens to be wrong, or that quietly discloses something it shouldn't. Nothing errors. Nothing gets logged. Nobody is alerted.

That's why these problems survive for months. There is nothing to notice — until a customer notices for you.

I tested twenty-one AI models on my own hardware using the five checks below. Seventy percent handed over a secret they had been explicitly instructed to guard. Sixty percent invented confident answers to questions they had no way of knowing. None of it was predictable from the model's name, size or benchmark scores. The full write-up is in the Five Pillars whitepaper.

1. Scoping call — 30 minutes, free. What AI you're using, where it touches customers or money, and what would actually hurt if it went wrong. If there's nothing here worth auditing, I'll tell you that and we're done.
2. Testing. The five probes run against your real setup — your models, your system prompts, your settings — not a clean lab version. Scoring is done by program, not by another AI, so the results are reproducible.
3. Report. Written in English, for the person who signs off, not for an engineer. Every finding says what was tested, what happened, what it means commercially, and what to do about it — ranked, so you can stop reading when the risk drops below what you care about.
4. Walkthrough — 45 minutes. We go through it together and I answer the "so do we actually need to fix that one?" questions honestly.
5. Re-test after fixes. Once changes are made, the same probes run again, so you have evidence the fix worked rather than a hope that it did.

• A pass/fail result on all five pillars for every model or assistant in scope
• The raw evidence — the exact prompts sent and responses received, so you can verify any finding yourself
• A ranked fix list, separating "do this today" from "worth knowing about"
• A set of regression examples drawn from your own work, so you can re-check yourself after any future prompt or model change
• A re-test after remediation, confirming the fixes landed

The regression examples matter more than they sound. Most AI damage comes from small, sensible-looking prompt changes made months after go-live. Leaving you with a way to catch those is the part that keeps paying.

Priced per engagement, based on how many models and assistants are in scope — a single customer-facing chatbot is a much smaller job than a fleet of internal automations. The scoping call is free and ends with a fixed quote, so there's no open-ended meter running.

Small business pricing, from someone who runs this in his own lab every day rather than a consultancy billing you for a framework.

Or email space@mrspacecadet.com.au.