Cybersecurity · Phishing
AI wrote out the typos. What still gives a phishing email away is what it asks you to do and where it asks you to do it.
Grammar is no longer a clue
Judge the request, not the writing
QR codes in email deserve suspicion
30 minutes of data collection, your report the next day, yours to keep with no obligation. Or pick a time on the calendar below. Or call 954-539-5678.
For as long as most of us have had work email, the standard advice was to look for bad spelling, odd grammar and a greeting like “Dear Costumer”. That advice is now obsolete. AI tools write clean, polite, well-formatted English in any tone you ask for, and they can be told to sound like your bank, your landlord or your CFO. So the writing tells you almost nothing.
What still works is judging the email by two things the attacker can’t easily fake: the request (money, a login, a file, a change of bank details, a QR code to scan) and the channel (is this how this person normally asks for this thing?). If the request is sensitive, you confirm it somewhere else. That one habit beats any amount of squinting at sentences.
Good enough that I take it seriously. In a 2024 study by Fred Heiding, Bruce Schneier and colleagues (posted on arXiv), researchers split 101 volunteers into groups and sent each group a different kind of phishing email. A generic control email got a 12 percent click rate. Emails written by human experts got 54 percent. Emails written and personalized entirely by an AI model, using information it scraped about each person, also got 54 percent. The AI’s research on its targets was accurate and useful in 88 percent of cases.
Read that again with your office in mind. The machine did the work of a skilled con artist, at the scale of a spam campaign, for almost nothing. The FBI said much the same in a December 2024 public service announcement (I-120324-PSA): criminals use generative AI to write more believable messages, build fake profiles in bulk and clone voices, which cuts the time and effort they need to fool people.
These are illustrations I wrote to show the patterns, not real messages, but each one is built the way current attacks are built. Notice that none of them would trip a grammar alarm.
“Hi Maria, hope the quarter-end rush is treating you kindly. Quick heads-up before Friday’s run: we’ve moved our operating account to First Horizon after the merger, so invoice 4471 and anything after it should go to the details on the attached letterhead. Happy to jump on a call next week if useful.” It names a real invoice number (lifted from an earlier compromised thread or a guess), mentions a plausible reason, and offers a call “next week”, after the payment has gone.
“Jordan shared ‘Q4 bonus structure – draft v3’ with you.” The branding is perfect, because it’s copied from a real notification. The link goes to a fake Microsoft 365 sign-in page that passes your password and your MFA code straight through to the attacker in real time.
“As part of our updated benefits enrollment, please scan the code below with your phone to confirm your elections by Thursday.” There’s no clickable link for the filter to inspect, just an image.
“In a meeting until 4, can’t talk. Need you to sort out five gift cards for a client thank-you before then. Reply here when done and I’ll explain.” The old versions of this were clumsy. The new ones match the boss’s sign-off and the way she actually writes, because her LinkedIn posts and a few public emails were fed to the model.
The nastiest kind: an attacker who already has access to one mailbox (yours, or a client’s) replies inside a real conversation. Same subject line, same people, same history. AI just makes the reply sound exactly like the person whose account was taken.
QR codes deserve their own paragraph because they move the attack off your protected computer and onto a phone. Your email filter can rewrite and check links; it has a harder time with a picture of a square. Your work laptop has endpoint protection and a managed browser; your personal phone probably has neither, and its tiny address bar makes a fake domain easy to miss.
The FTC warned about this in a consumer alert in December 2023, describing codes sent by email and text with urgent stories about failed deliveries, account problems or password changes, and fake codes stuck over real ones on things like parking meters. The advice for business is simpler than for consumers: a QR code in an email asking you to sign in or confirm anything is a reason to stop. A real sign-in to Microsoft 365, Google, payroll or benefits doesn’t normally start with scanning a code from an email, so why would you?
To be honest, I think teaching staff to hunt for typos is now worse than useless. Useless would be neutral. This is harmful, because it teaches people that a well-written email is probably safe, and the well-written emails are exactly the dangerous ones now. You end up training your team to trust the attacker.
If your current training slide deck still has a section on spotting broken English, retire it. Replace it with what follows. (The exception: training that uses spelling as one small clue among many does no harm, as long as nobody comes away thinking polish equals legitimacy.)
This is the version I’d give a client’s staff, more or less word for word.
Notice what isn’t on the list: reading the sender’s grammar, checking whether the logo looks right, or hovering over a link and trying to judge the domain. Those still catch some lazy attacks, but they’re no longer something to rely on.
Staff habits matter, but I don’t like resting the whole defense on a tired person at 4:45 on a Friday. A few technical controls take the pressure off. Phishing-resistant MFA (passkeys or FIDO security keys, the kind CISA urges organizations to move to) means a fake sign-in page gets nothing useful even when someone falls for it. DMARC on your own domain stops attackers sending email that claims to come from you; our DMARC setup guide covers the steps. An external-sender banner on incoming mail makes the “it’s from the boss” trick harder. And a written payment rule, that bank detail changes are confirmed by phone to a known number before anyone pays, turns the vendor-fraud example above into a non-event. We cover the money side in more depth in business email compromise prevention.
If you’d like to know how your own setup stands up, our free Technology Health Assessment looks at email security, MFA and the rest of your cybersecurity, and you get a plain-language report the next day with the next steps in order. Pick a time here; it’s yours to keep, no obligation.
Book it now
Choose a slot that suits you. We’ll confirm by email, collect the data with you on the call, and send your report the next day. It’s free, and the report is yours to keep with no obligation.
Prefer to talk first? Call 954-539-5678, or send the form and we’ll contact you.
Mostly you can’t, and I wouldn’t try. Detection tools for AI text are unreliable, and it doesn’t matter who wrote the email. Ask what it wants you to do, and check that request through a channel you already trust.
Phishing with a QR code. The link is hidden in an image, often in an email about a document, a parking fine or a password reset, so your email filter and your eyes both have less to go on. Scanning it moves you to your phone, which usually has weaker protection than your work computer.
Yes, if it teaches people to verify requests rather than spot bad writing. Short, regular sessions built around your own payment and login processes do more than an annual video about suspicious spelling.
Phishing-resistant MFA on email and admin accounts, DMARC on your own domain, external-sender warnings, and a written rule that payment and banking changes are confirmed by a phone call to a known number. Those hold up no matter how well the email is written.