Lock the Doors: AI Security Basics for a Small Office · Lesson 6 of 7

Fakes, cloned voices and the call-back rule

20 minutes. One rule, written down, that beats the best-written fake.

---

Everything so far has been about locking doors: updates, permissions, connections. This lesson is about the attack that doesn't need a door at all. Someone just asks, convincingly, and a helpful person on your team says yes.

It's an old trick. A fake invoice from a vendor. An email that looks like it's from the owner, asking for gift cards. A "new bank details" letter from a supplier. What changed is that AI makes the fake much better and much cheaper.

The writing got good. The tell-tale signs people were taught to look for (bad grammar, odd phrasing, a strange tone) are mostly gone. AI writes a clean, polite, on-brand email in any style, in seconds.

Official-sounding names are cheap to fake. In October 2026, OpenAI banned accounts tied to Russian and Iranian influence operations. The Russian accounts' own reports described a fake think tank and posing as a regional education office to get schools to run events. The Iranian accounts invented seven fake authors and placed almost 100 articles with small media outlets. None of it needed hacking. Just convincing writing under a borrowed name. That's exactly the trick used against small businesses in fake vendor and fake official emails.

Voices can be cloned. The FBI warned in December 2024 that criminals use AI to generate short audio clips in a loved one's voice, asking for money in a crisis. The same tool works for a boss's voice, or a vendor's. A familiar voice on the phone is no longer proof of who's calling.

You can't train your team to spot every fake. The fakes are too good now. So don't try. Make the fake not matter.

Here it is in one sentence:

> Before we move money, change payment details, or send private data because someone asked, we confirm by calling them back on a phone number we already had.

The key words:

• "Call them back." Not reply to the email. Not answer the text. You start a new conversation, on a channel the requester didn't choose.
• "A number we already had." From your own contacts, an old invoice, the vendor's contract or the official website you looked up yourself. Never the number in the email, the signature or the voicemail. A faker puts their own number there.
• Every time, no matter who it seems to be from. Especially if it's the owner. If the owner is the one asking, the owner should be glad you called.

This matches the FBI's own advice for voice cloning: hang up, look up the contact for the person or organization yourself, and call that number directly.

Use the rule for any request, by email, text, phone, chat or letter, that asks you to:

• Pay something new, or pay something faster than usual
• Change bank details, a payment address or a payroll deposit
• Buy gift cards, or send money in any unusual way
• Send customer, employee or financial records
• Share a password, a code, or reset someone's sign-in
• Make an exception "just this once" because it's urgent or confidential

And use it on anything that comes from an official-sounding source you've never dealt with: an "authority," an "institute," a "regional office," a lawyer you haven't hired.

Pressure is the signal. Urgency ("today"), secrecy ("don't tell anyone yet"), and authority ("this is the owner," "this is the agency") are the three levers every fake pulls. When you feel all three, slow down and call.

• A code word. The FBI suggests families agree on a secret word or phrase to confirm identity. A small team can do the same for the one or two people who can approve payments. If the "owner" calls asking for a transfer and can't give the word, it isn't the owner.
• Two people for money changes. In a small office, that can be as simple as: whoever receives a request to change bank details doesn't make the change. They hand it to someone else, who makes the call-back.

Your AI assistant can help here, carefully:

• Paste a suspicious email into your assistant and ask: "List anything in this email that asks for money, data or a change in how we pay someone, and any signs of pressure." It's a good second look. It is not a verdict. Even if the AI says it looks fine, the call-back rule still applies.
• If your AI is connected to your email (Lesson 4), it should never be allowed to pay, change payment details or send records on its own. A fake email that tricks a person can trick an AI too.

Open the Call-Back Rule Card template.

• Fill in who approves payments and payment changes (one or two names).
• Fill in where you look up a trusted number for your main vendors, your bank and your key customers (your contacts app, the vendor list in your accounting software, past invoices).
• Decide whether you want a code word for payment approvals. If yes, agree on it in person, and don't write it on the card.
• Print the card or post it where your team works, and tell them about it in person. Say this part out loud: "You'll never be in trouble for calling to check. You might be for not calling."

Check it worked

• The card names who approves money changes and where trusted numbers come from.
• Everyone who handles payments, payroll or customer records has seen it and heard it from you.
• You can describe the rule in one sentence without looking.
• Bonus: next time a real vendor emails about a payment change, you call them back, and they're glad you did.

---

Next: put it all together into one short weekly habit.

→ Lesson 7: Capstone: your 10-minute weekly check

Previous: Find the AI tools nobody told you about · Next: Capstone: your 10-minute weekly check