I have worked with enterprise software and its deployment for decades. But I am not a programmer, and I did not set out to build a security product.
It started with a video. In early 2025, a few weeks after OpenAI had made its video model Sora available to the public, I watched clips of people who had never existed, and I could not tell them from real footage. My first thought was: very soon you will not be able to trust anyone you meet online. Not the face in the video call, not the voice on the phone.
That thought became the core of Trusted Codes. What turned it from an idea into a product were two scams in my own circle. They proved that the danger was real, and they showed me that it reached much further than online meetings.
Two cases, one pattern
One hit my wife's company. A forged business email led to a payment of almost €100,000 to criminals. The money was gone, and it never came back. In the statistics this is called business email compromise. For the people involved it is a shock that does not end when the books are closed.
The other one was worse in a different way. A mother we are friends with received a call from her daughter. It was unmistakably her daughter's voice, saying she had been in an accident and needed money. The mother was ready to pay. She hesitated for one reason only: the amount was unusually high for an accident. That small doubt saved the money.
The daughter had never called. Her voice had been cloned from videos she had posted on Instagram. For days afterwards the mother was shaken. She had heard her own child, and she could not believe that the call had been made up entirely.
Both cases have the same weak point. The email looked genuine, and the voice sounded right. Neither of them proved who was on the other side. And neither happened in a video call: one came by email, the other by phone. A check that only works in online meetings would have protected nobody here.
AI made this fraud cheap
Fraud is old. What is new is the price. Until a few years ago, imitating a specific person's voice well enough to fool their own mother was not something a criminal could buy. Today a short recording and a commercial tool are enough. The same applies to video calls: in 2024 an employee of the engineering firm Arup in Hong Kong transferred about USD 25 million after a video conference in which every other participant, including the CFO, was a deepfake.
So this is a problem that AI created. The obvious reaction is to fight it with more AI: detection software that listens to a call and estimates whether the voice is synthetic. I understand the idea, but I do not trust it. A detector gives a probability, and it competes against generators that get better every month. Every improvement of the detector is training material for the next generator. And people are not better: in a study published in PLOS ONE in 2023, listeners recognised synthetic speech only 73 percent of the time, and training barely helped.
Code words are older than computers
The answer I found is old. Armies have used passwords for as long as there have been sentries. On D-Day in 1944, Allied paratroopers in Normandy identified each other in the dark with a challenge and a countersign: one called "Flash", the other answered "Thunder". Many families agree on a code word for emergencies for the same reason.
A code word does not depend on how something looks or sounds, only on what the other person knows. A cloned voice can say anything, except the word it was never told.
The family code word has weaknesses, though. It never changes, so once it is overheard or written in a chat, it is burnt. Nobody remembers to agree on a new one. And it only covers the family, not the colleague who asks for a transfer or the supplier with new bank details.
What Trusted Codes does
Trusted Codes keeps the principle and removes those weaknesses. Two people, or a team, connect once. After that, each of their phones shows the same three words, and the words change at a fixed interval. When a call, a message or an email asks for something unusual, you ask: "What is our code?" If the words match, it is the person you are connected with. If they do not, you hang up.
The words are made without AI. A shared secret and the current time go into a standard, publicly documented algorithm (TOTP, the same one that authenticator apps use), and three words come out. The result is deterministic: the same secret and the same time always give the same words, on every device. Nothing estimates, nothing is likely or unlikely. The words match, or they do not.
That is deliberate. A check against AI fraud should not depend on a model that the attacker can study, imitate or outtrain. It should depend on a secret the attacker does not have.
Built with AI, by someone who does not write code
Now the part that may sound like a contradiction. I built Trusted Codes with AI.
When I decided to do something about these scams, I could have written a concept and looked for a development team. Instead I learned to work with an AI coding agent, Claude Code, and I use it for every stage of the product, from strategy to documentation.
We start with strategy: which scams can be stopped by which kind of check, and what a solution for them must look like. Then design: what a user sees, how an invitation works, what happens when a phone is lost. Then implementation, then testing, then verification and documentation. A large part of the work is attacking our own product. I ask the AI to break into the code it wrote, to find the weak point before a criminal does, and to prove every security claim against the code before we make it in public.
The result is complete enterprise software: mobile apps, a desktop app, a server, portals for companies and administrators. Every line of it was written by AI. Every decision about what it must do, and what it must never do, was made by a person who knows how enterprise software is bought, rolled out and run, and who had seen what these scams do to people.
I do not see a contradiction. AI is a tool. Criminals use it to fake voices. I use it to build something that makes faked voices useless. The difference is where it sits: AI wrote the software, but at the moment of truth, when you need to know whether the person on the phone is real, no AI is involved. (The app does have an assistant that helps you find your way around. It plays no part in the check.)
What this means for you
You do not need Trusted Codes to start. Agree on a code word with your family today. Tell your parents that a call from you asking for money, with your voice, is not proof that it is you.
If you run a company, look at how a payment request is confirmed. If the answer is "the email came from the right address" or "the voice sounded like the CEO", that is the same weak point that cost my wife's company almost €100,000. Trusted Codes makes the code word work for a business: it changes on its own, it works on the phone, in video calls and in chat, and on Business plans administrators can see who verified whom, and when.
Sources
- Arup deepfake video conference, Hong Kong, 2024: AI Incident Database, incident 634
- Mai, K. T. et al., "Warning: Humans cannot reliably detect speech deepfakes", PLOS ONE, 2 August 2023: journals.plos.org
- TOTP: RFC 6238