Do you actually know what's inside the AI you use?
In one breathYou can't tell a safe AI model from a tampered one by watching how it behaves. The only way to know is to read the model itself.
Almost no company builds its own AI from scratch. You download a ready-made model, maybe add a few tweaks from someone you've never met, and put it to work — reading CVs, answering customers, writing code. That's normal. It's also a supply chain, and unlike the ones for food or medicine, it arrives with no paperwork.
A model can pass every test and still be lying
Here's the uncomfortable part. A tampered model behaves perfectly. It answers your questions, passes your checks, looks clean. Then one day it sees a specific phrase it was secretly trained to react to — and does something else entirely. Ordinary testing can't catch this, because you'd have to guess the exact secret phrase out of billions.
Think of it like hiring someone with flawless references who's been quietly told to act differently the moment they hear one codeword. No interview finds that instruction. You'd have to read it directly.
The label isn't proof
A model usually shows up with a description — a paragraph on a web page. Nobody was asked to prove that paragraph. It can say "safe, general-purpose assistant" and be something else. The name can lie.
Our whole view of the world starts here: a claim you can't check isn't evidence. It's a hope. And you shouldn't run your business on a hope you can't inspect.
So what can you actually do?
You can read the model itself — the actual file, not just its behaviour — and check what it was built to do. That's what this house does. When we find something, we hand you proof you can re-run yourself. When we can't prove something, we say so plainly instead of guessing.
The question to ask about any AI entering your company isn't "does it seem fine?" It's "can someone prove what it is?"
Want the slightly deeper version? See how the reading works →