Skynet didn't show up with chrome skeletons and laser rifles. It showed up as a text from your bank, a call from your kid, and a helpful browser extension with five stars.
I spend my days locking down systems for a living, and the honest truth is that most of what keeps you safe from AI-powered scams isn't technical. It's a handful of habits. Here are the ten I actually use and the ones I've drilled into my own family. None of them require you to understand how a language model works.
1. The 600 series had rubber skin
In The Terminator, Kyle Reese explains that the old infiltrators were easy to spot because of their rubber skin. The new ones look human. Phishing just made the same upgrade.
For twenty years the advice was "look for typos and bad grammar." Retire it. The FBI warned in late 2024 that criminals use generative AI specifically to clean up the spelling and grammar mistakes that used to give them away. 1Password puts it bluntly: don't rely on obvious mistakes to spot a scam.
What still gives a scam away is the ask. Urgency, secrecy, money, a login, a code, a gift card. The polish is free now. The pressure is the tell.
Judge the request, not the writing.
2. Ask about Wolfie
The best security control in Terminator 2 is a dog's name. John calls home, and the T-800 grabs the phone, imitates John's voice, and asks his "foster mom" what's wrong with Wolfie. The dog's name is Max. The T-1000 says Wolfie's fine, and the T-800 hangs up: "Your foster parents are dead."
Voice cloning is real now. The FBI describes criminals generating short clips of a loved one's voice in a fake crisis, asking for money or a ransom. Its first recommendation is the Wolfie trick: create a secret word or phrase with your family. If you go that route, pick a word that isn't on anyone's Facebook, and agree that anyone calling in a panic for money gets asked for it.
Full disclosure: my family doesn't use one. I just hang up and call the person back directly to confirm, which is rule 3. Either works. What doesn't work is deciding on the fly, while someone you love is crying on the phone.
A machine can copy a voice. It can't copy what it never heard.
3. Hang up and call back
In early 2024 a finance employee at the engineering firm Arup got an email he suspected was phishing. Then he joined a video call where the CFO and several colleagues looked and sounded right. None of them were real. He sent about $25.6 million across 15 transfers.
The call didn't beat his suspicion. It talked him out of it. The FBI's advice works on video calls and voice calls alike: hang up, look up the real number yourself, and call back. Not the number in the email. Not the number the caller gives you.
Most of us won't get a $25 million boardroom deepfake. The machine that calls my house is dumber and more persistent: AI robocallers "from collections," chasing medical bills that you know for a fact don't exist. You can't reliably tell the bot from a person anymore, so don't try. Scam mailers for fake debts have been around for years. What changed is that a scammer can now point an AI at a list, spoof a believable caller ID, and let it run. Spoofing caller ID to defraud someone is already illegal under the Truth in Caller ID Act, and the FCC ruled in February 2024 that AI-generated voices count as "artificial" under the robocall law, which makes them illegal in robocalls without your consent. None of that stops the calls. It just tells you who's breaking the law.
The FTC's red flags for a fake debt collector are a debt you don't recognize, a caller who won't give a mailing address or phone number, and threats of arrest or law enforcement. A real collector has to give you "validation information" in writing or over the phone: who they are, who the original creditor is, and exactly how much you supposedly owe. If you still don't recognize the debt, you can dispute it in writing within 30 days.
The best defense is boring: keep track of your own bills. If you know what you owe and to whom, a robocall about a mystery medical debt is just noise. Don't pay, don't confirm your name or birthday, and don't press 1 to talk to anyone. Hang up. If you're not sure, call the billing office at the number on a bill you already have.
If it's real, it'll still be real in five minutes.
4. Look again
In The Matrix, Morpheus walks Neo through a crowd and asks if he was listening, or looking at the woman in the red dress. "Look again." She's Agent Smith with a gun.
That's every fake login page. It looks right, so you don't look at the address bar. Here's the cheat code: a password manager reads the domain so you don't have to. 1Password, for example, won't autofill when the site's address doesn't match the one you saved, and it now warns you if you try to paste the password in anyway.
At home we run a self-hosted Bitwarden instance for the whole family. Everyone has their own login, and the things we actually need to share, passwords and shared information, live in a shared organization that each person can reach from their own vault. Bitwarden calls these organizations and collections. The same setup works with 1Password or LastPass. Nobody texts a streaming password around, and nobody types one in by hand.
One honest caveat: some legitimate sites just don't autofill properly. When you save credentials for one of those, make a note of it in the entry. Then be extremely careful every time you log in there, and check the domain yourself, character by character, because that's exactly the gap a lookalike page is counting on.
So when your password manager refuses to fill on a site that normally works, that's not a bug. That's the woman in the red dress turning around.
If autofill won't fill, don't paste. Look again.
5. Passkeys beat the man in the middle
Modern phishing kits don't just steal your password. The Center for Internet Security lays out how these adversary-in-the-middle attacks work: a reverse proxy sits between you and the real site and passes everything through. You complete your text code or app approval for real, and the attacker keeps the logged-in session cookie for themselves. Your MFA worked. It just worked for them.
Passkeys fix this. A passkey is tied to the real website, so it won't hand itself over to a lookalike sitting in the middle. The FIDO Alliance, which writes the standard, says a passkey on its own is stronger than a password plus a text code or app prompt, because those are still phishable. Turn on passkeys anywhere they're offered, starting with email, your bank, and your phone carrier.
My own preference is a YubiKey plus Bitwarden. A good master password is fine, but a hardware key guarding both the password vault and your passkeys is glorious. I highly recommend the biometric YubiKeys: the Bio series uses your enrolled fingerprint to approve a passkey sign-in, with the key's PIN as a fallback. (I covered my own key setup in Part 2 of the Radar series.)
Non-biometric keys use a PIN. Never write it down, save it, store it, or share it. And if you're entering it in public, cover the keypad with your spare hand, because cameras are everywhere and AI is very good at reading hands. Researchers trained a model on video of people typing ATM PINs while covering the keypad with their other hand, and it still guessed 30% of 5-digit PINs within three tries. Their test of shielding found it only really worked when the whole keypad was covered. So cover all of it.
A text code can be relayed. A passkey can't be fooled about where it is.
6. Don't take phone numbers from the machine
A man searched for Facebook support, found a sketchy number, and asked Meta's own AI chatbot whether it was legit. The chatbot said yes. It had made that up. The number belonged to a scammer, who talked him into installing an app. It's also been reported that scammers are seeding fake support numbers across the web, hoping they get scooped up into AI search answers.
An AI answer sounds sure of itself whether it's right or wrong. Get support numbers from the back of your card, your bill, or the company's own site that you typed in yourself.
The machine is confident. Confidence isn't verification.
7. Anyone you haven't unplugged is potentially an agent
Morpheus's warning about the Matrix applies to your browser. In late 2025, OX Security found two Chrome extensions posing as a popular AI sidebar, with 600,000 and 300,000 users. The bigger one carried Google's "Featured" badge. They asked permission to collect "anonymous" analytics, then shipped people's entire ChatGPT and DeepSeek conversations, plus every website they visited, to the attackers every 30 minutes.
"Featured" isn't vetted. Five stars isn't safe. An extension that can "read and change all your data on all websites" can read everything you type into a chatbot.
I've forgotten about plenty of extensions and apps over the years. What saves me is that I reinstall my entire desktop from scratch once a year, and anything I didn't miss by then simply doesn't come back. You don't have to go that far, but go to your browser's extensions page today and remove anything you don't use. Better yet, be picky about what you install in the first place. That's true of all security, not just the AI kind.
Every extension is a stranger reading over your shoulder.
8. The AI that reads the page can be read by the page
AI browsers that can click and type for you are the new shiny thing, and the attack is ugly in its simplicity. Brave's researchers hid instructions in a Reddit comment. When a user of Perplexity's Comet browser clicked "Summarize this page," the AI followed the hidden orders. It found the user's email, triggered a login code, read the code out of their Gmail, and posted both back to Reddit. The user clicked one button.
The AI couldn't tell the page's words from the user's words. Brave's own advice is to treat page content as untrusted and keep agent-style browsing separate from the browser where you're logged into everything.
So treat an AI browser as insecure. Use it for research, not for anything that needs a login. Don't bank in it, don't leave it signed into your email, and never reuse a password or use an easy one there. If it never holds a credential, a hidden instruction has nothing to steal.
It's not the summary that's dangerous. It's who the summary is taking orders from.
9. Don't tell the machine your secrets
Whatever you type into a chatbot lives somewhere: on the company's servers, in your history, and potentially in a shared link. There have been reports of shared chatbot conversations, some with names and personal details, turning up in search engine results. Add the extensions from rule 7, and your chats can leak in ways you never agreed to.
I'll confess this one. When I was first learning to work with AI, I put API tokens and similar secrets into base64-encoded files and gave the AI access to them, so it could wire them in wherever they were needed across multiple systems. Base64 isn't protection. The standard that defines it says outright that it visually hides things like passwords but provides no real confidentiality. Anything that can read the file can decode it.
I was never bitten. Those were dev builds, the credentials lasted a couple of hours, and by the time anything went to production it used a real secrets manager and/or Kubernetes secrets. But my dev systems were technically exposed and vulnerable every time I did it. Short-lived credentials limited the damage. They didn't make it safe.
For everyone else, the rule is simpler: don't paste in account numbers, Social Security numbers, passwords, medical records, or anything you'd hate to see on a billboard. Check your chatbot's settings for chat history and training, and delete shared links you don't need.
Assume every chat could end up in front of a stranger, and type accordingly.
10. Shrink your digital double
You can't be cloned from footage that doesn't exist. The FBI recommends limiting the photos and voice recordings of yourself you post publicly, making social accounts private, and limiting followers to people you actually know. The Arup deepfakes were reportedly built from publicly available video and audio of real staff.
You don't have to vanish. Just stop handing out free training data.
Deja Vu
In The Matrix, Trinity says that a Deja Vu is usually a glitch, and that it happens when they change something. Treat the strange moments the same way. The voicemail that's almost your kid's voice. The login page your password manager won't fill. The support number from an AI that you can't find on the company's site. Those are your black cats.
The habits behind all ten rules are simple. Judge the ask, not the grammar. Verify on a channel you picked. Let passkeys and your password manager check the domain for you. Keep AI tools away from your money and your secrets. Kyle Reese said the Terminator couldn't be bargained with or reasoned with. Neither can a scam. You don't have to win the argument. You just have to hang up.
Gandalf summarized the whole playbook back in 1954: "Keep it safe, keep it secret.", because the machine won't do it for you.