AI-Powered Social Engineering: Deepfakes, Voice Clones and AI Phishing

Generative AI did not invent a new kind of scam. It made the old ones cheaper, cleaner and more personal. The psychological levers have not changed: authority, urgency, fear, helpfulness and secrecy. What changed is that a criminal can now write fluent email in any language, clone a voice from a short clip, and put a convincing face on a video call.

That means the old tells (bad grammar, a strange accent, a blurry photo) no longer protect you. The defence that still works is verification that does not depend on how real the request looks. Call back on a number you already had. Ask a question only the real person can answer. Get a second person to approve. Never move money or credentials on the same channel the request arrived on.

This guide covers each tactic, the habit that defeats it, and what to do if one gets through. The company payment controls for executive impersonation live in Deepfake Executive Fraud Controls.

What generative AI changed, and what it did not

The FBI put the shift plainly in a December 2024 public service announcement: "Generative AI reduces the time and effort criminals must expend to deceive their targets" (FBI IC3, I-120324-PSA). The same notice lists AI-written messages, fake profile photos, forged ID documents, cloned voices of relatives and fake real-time video calls with executives.

The UK's National Cyber Security Centre reached a similar judgement in January 2024: AI "provides capability uplift in reconnaissance and social engineering" (NCSC, The near-term impact of AI on the cyber threat). Europol's 2025 internet crime assessment describes generative AI being used to "supercharge social engineering attacks", with messages tailored to a victim's culture and personal details (European Commission summary of Europol IOCTA 2025).

What is new

• Scale. One operator can run many personalised conversations at once.
• Polish. Spelling, grammar and tone are no longer reliable signals.
• Personalisation. Public posts and org charts feed messages that name your real manager, project or family.
• Media. Voice and video, once the "safe" channels, can now be faked.

What is not new

• The request still asks you to act fast, keep it quiet, or skip a normal step.
• The request still ends in money, gift cards, crypto, a code, a password, a remote-access session or a file you run.
• The attacker still does not control the phone number in your contacts, your company directory, or the second approver down the hall.

You will not out-spot a good fake. You can make it irrelevant.

Common mistakes

• Training people to look for typos. That lesson now teaches people to trust clean writing.
• Assuming a phone call or video is more trustworthy than email.

Action checklist

• Stop using spelling or grammar as a reason to trust or distrust a message.
• Write down the three requests you would never act on without a callback: money, credentials, and remote access.
• Treat urgency plus secrecy in the same request as a red flag on any channel.
• Review what you post publicly about your family, travel and job role.

AI-written phishing: grammar is no longer the tell

For two decades, awareness training taught people to spot phishing by its mistakes. The NCSC notes that generative AI can produce lures "without the translation, spelling and grammatical mistakes that often reveal phishing" (NCSC). The FBI says criminals use AI to translate messages and hide grammar errors (FBI IC3).

Spear phishing (a message tailored to one person) used to take research time. A model now turns your public profile into a believable note from a "recruiter" or your "CFO" in seconds.

Tells that still work

• The ask, not the prose. Does it want a payment, a login, a code, a document signed or a file opened? Judge the request, not the writing.
• The channel switch. "Let's move to WhatsApp" or "text me on my personal number" is a common step. The FBI flagged exactly this pattern in a 2025 campaign that used AI voice messages to impersonate senior US officials (FBI IC3, I-051525-PSA).
• The sender and link domain. Check the actual domain, not the display name.
• The pressure. Deadlines, threats of account closure, or "I'm in a meeting, just do this" are design features of the scam.

When in doubt, do not reply or use any contact details in the message. Open the service from a bookmark, reach colleagues through your company directory, and report the message.

Stolen passwords are what most phishing is after. A passkey cannot be typed into a fake page, which removes the prize. See credential hygiene for the switch, and why SMS codes are not real MFA.

Common mistakes

• Clicking "just to check" whether a link is real. Go to the site yourself instead.
• Trusting a blue check or a verified badge. A badge proves an account paid or registered, not that the request is legitimate. See Blue Checks Are Not Security.

Action checklist

• For every unexpected request, name the ask (money, login, code, file) before you read the rest.
• Bookmark the login pages for your bank, email and payroll so you never need a link.
• Learn where your organisation's report-phishing button or address is.
• Turn on passkeys for your main email account where the provider offers them.
• Refuse any request to continue a work conversation on a personal messaging app.

Voice clones: family emergencies and executive calls

A cloned voice is a synthetic copy of someone's speech, generated from a recording. The FTC warned in 2023 that a scammer needs only "a short audio clip of your family member's voice" plus a voice-cloning program, and that the clip can come from content posted online (FTC consumer alert, March 2023). Microsoft Research's VALL-E X project describes speech synthesis from "a 3-second enrolled recording of an unseen speaker", and itself names "impersonating a specific speaker" as a misuse risk (Microsoft Research, VALL-E X).

Plan as though your voice, and your family's, can be copied.

The two common scripts

• Family emergency. A panicked call from a "grandchild", "son" or "daughter": an accident, an arrest, a kidnapping. Someone else (a "lawyer", "officer") takes over and asks for bail money, often by wire, crypto or gift cards. See the Scam Atlas entry on the grandparent voice clone scam.
• Executive call or voicemail. A "CEO" or "CFO" asks finance to make an urgent, confidential payment, or asks IT to reset an account. See deepfake CEO voice fraud and help desk reset impersonation.

What defeats it

1. Hang up and call back. The FTC's advice is direct: "Don't trust the voice. Call the person who supposedly contacted you and verify the story" (FTC). Use a number you already had, not one given on the call.
2. Agree a safe word. The FBI recommends you "create a secret word or phrase with your family to verify their identity" (FBI IC3). Pick something never posted online, share it in person, and agree the rule: no word, no money.
3. Try another route. If you cannot reach the person, call their partner, roommate, school or office.
4. Watch the payment method. Requests for wires, crypto or gift card numbers are a warning sign because the money is hard to recover (FTC).

Common mistakes

• Using voice recognition as proof. "It sounded exactly like her" is the scam working as designed.
• Using voice biometrics as the only check in a bank or help desk flow. A control that authenticates on voice alone is now a target.

Action checklist

• Agree a family safe word in person this week and tell every adult in the household.
• Save direct numbers for close family, plus one backup contact for each.
• Write a one-line rule on the fridge: "Crisis call asking for money? Hang up and call back."
• Set social accounts that carry your voice or your family's voice to private where you can.
• If your bank offers voice ID as the only check, ask for an alternative.

Deepfake video calls

A deepfake is synthetic video or audio that makes a person appear to say or do something they did not. In early 2024, Hong Kong police described a finance worker at a multinational who joined a video call with what appeared to be the CFO and other colleagues. All were deepfake re-creations. The worker sent HK$200 million (about US$25.6 million) across 15 transactions (CNN, May 2024).

The engineering firm Arup later confirmed it was the company and said "we can confirm that fake voices and images were used". It added that "none of our internal systems were compromised" (same report). Two details matter for you:

• The employee was suspicious at first. The opening message looked like phishing because it asked for a secret transaction. The video call is what removed the doubt.
• Nothing was hacked. The money moved because one person could act on the strength of a conversation.

The Scam Atlas entry on the deepfake CFO video call tracks this pattern. The longer analysis is in The $25 million deepfake: why your video calls can no longer be trusted.

What defeats it

• Treat the call as a request, not as proof. A face on screen confirms nothing about who is behind it.
• Verify outside the call. Message the person on your company chat, or ring their directory number, and ask them to confirm.
• Insist on the normal process. If the payment needs two approvers or a ticket on any other day, it needs them today.
• Change the plan. Ask the person to hang up and call you back on their known number. Real colleagues can do this. A rehearsed fake often cannot.

The FBI notes that fakes can show "subtle imperfections", such as odd hands, teeth or eyes, lag, or unnatural movement (FBI IC3). Use these as a reason to slow down. Never use their absence as a reason to trust.

Common mistakes

• Thinking a multi-person call is safer. In the Hong Kong case, several "colleagues" made it more convincing.
• Accepting "I'm travelling, I can't use the normal channel". That excuse is part of the script.

Action checklist

• Decide now that no video call, however real, is enough to approve a payment or access change on its own.
• Save your manager's and finance contacts' directory numbers where you can find them in 30 seconds.
• Ask your finance team whether a "confidential" payment request has its own written process.
• Practise the line: "I'll call you back on your desk number to confirm."

Fake recruiters, fake interviews and fake IT workers

Job seekers are targeted with fake offers, and employers with fake candidates.

When you are the candidate

The FBI described North Korean actors who research employees through professional networking sites and then pose as recruiters or tech companies, with convincing websites and stolen photos (FBI IC3, I-090324-PSA). A key warning sign is a "pre-employment test" or debugging exercise that asks you to run unfamiliar packages, scripts or code repositories. The FBI advises against running such code on company devices.

The FTC adds the classic job-scam signs: you are asked to pay for training or equipment, or sent a check and told to send part of it back (FTC, Job Scams). Its rule is simple: "Honest employers, including the federal government, will never ask you to pay to get a job."

See the Scam Atlas entries on the fake recruiter and fake interview malware.

If a test genuinely requires running code, the FBI suggests a virtual machine on a non-company device.

When you are the interviewer

The FBI warned in 2022 of deepfakes and stolen identities used to apply for remote IT roles, noting lip movement that does not "completely coordinate with the audio" (FBI IC3, I-062822-PSA). In January 2025 it reported North Korean IT workers "using artificial intelligence and face-swapping technology during video job interviews" (FBI IC3, I-012325-PSA).

The FBI's recommendations for employers include identity checks that continue through onboarding, cross-checking applicants for reused resumes and phone numbers, and to "complete as much of the hiring and onboarding process as possible in person" (FBI IC3, I-012325-PSA).

Common mistakes

• Treating a polished recruiter profile as proof. AI makes profiles and photos cheap.
• Letting the person who passed the interview be assumed to be the person doing the work. Verify again at onboarding.

Action checklist

• Look up any recruiter's company independently before replying.
• Refuse to run interview code or install interview software on a work device.
• Walk away from any job that asks you to pay, reship goods, or forward part of a check.
• If you hire, ask HR whether remote candidates are re-verified at onboarding.
• Report suspected fake candidates to your security team, and in the US to IC3.

Fake support, chatbots and callback scams

Fake support flips the usual flow. Instead of the attacker calling you, you are tricked into calling them, or into accepting their "help".

• Callback phishing. An email (a fake invoice or renewal) gives a phone number to "cancel". The person who answers talks you into installing remote-access software. See callback phishing.
• Tech support pop-ups. A browser page claims your computer is infected and shows a number. The FTC is clear: "Real security pop-up warnings and messages will never ask you to call a phone number" (FTC, Tech Support Scams). See tech support pop-up.
• Email bombing then "IT support". Microsoft described a group that floods a target's inbox with subscription sign-ups, then calls posing as IT support to "fix the spam". It then gets the target to open Quick Assist (Windows remote help) and grant control. That access led to ransomware (Microsoft Threat Intelligence, May 2024). Microsoft later reported the same group contacting targets through Microsoft Teams. See email bombing and fake IT support.
• AI chatbots on scam sites. The FBI notes criminals embed AI chatbots in fraudulent investment and crypto sites to steer victims toward malicious links (FBI IC3).

What defeats it

• Only contact support through the number or chat on the company's official website or app, which you open yourself.
• Never let anyone remote into your device unless you started the contact through a channel you trust. Microsoft's guidance is the same: "Don't provide access to anyone claiming to have an urgent need to access your device."
• At work, check for the "External" label on Teams or chat messages from "IT". Microsoft recommends training users to look for it and never to approve sign-in requests over chat.
• Never read out a verification code or approve a login prompt you did not start. See MFA fatigue.

Common mistakes

• Calling the number in an invoice email to dispute a charge. Check your actual account instead.
• Assuming a chatbot is the company's because it appears on a site that looks like the company's.

Action checklist

• Save your IT help desk's real number and chat channel today.
• Close any pop-up that asks you to call a number. If the browser is stuck, force-quit it.
• Never install remote-access software because someone called you.
• Check the "External" label before trusting a Teams or chat message from "IT".
• Check suspicious charges by logging into the account directly, never by calling the number in the email.

Verification habits that work however good the fake is

These habits work because they never ask you to judge whether the voice, face or text is real. They move the decision to a channel the attacker does not control.

The five habits

1. Callback on a known number. Use your contacts, your company directory, or the back of your card.
2. Safe words and challenge questions. One for your family. For finance teams, a challenge agreed in person or tied to an internal system.
3. Second-person approval. Any payment, payment-detail change or access reset gets a second, independent approver. Two approvals gathered in one chat thread is one approval.
4. Slow down urgent requests. Urgency is the attack. A real emergency survives a five-minute callback.
5. Never one channel for money or credentials. Confirm a phone request by chat, an email request by phone.

Team habits hold only if leaders back them in advance. The company version, with thresholds, vendor-change delays and a blameless escape hatch, is in Deepfake Executive Fraud Controls. Day-to-day workplace routines are in Workplace Security Habits. Front desks face their own variant: see AI receptionist social engineering defense.

What detection tools can and cannot do

Deepfake detection tools look for signs of synthesis. They can help in places you control the capture, such as identity verification with a liveness check. They are weaker on a third-party video call or a phone line, where the decision sits with a human under time pressure. Treat a "likely real" score as one input, never as approval. If your organisation is evaluating products, the deepfake detection tools comparison sets out criteria.

Provenance is a different approach. Content Credentials based on C2PA attach signed history to media at creation, so you can check where it came from. The Content Credentials guide explains how. Neither detection nor provenance replaces a callback. For background on how fakes are made, see Deepfake 101.

Common mistakes

• Making the callback optional "if you are unsure". Make it automatic.
• Picking a safe word that appears in a social post or a birthday card photo.

Action checklist

• Print or save the channel table above where your household or team can see it.
• Agree a challenge question with your manager and finance contacts for payment requests.
• Ask that every payment-detail change at work requires a second approver.
• Practise saying "I'll call you back" to an urgent request, even a real one.
• Never approve money or credentials on the channel the request arrived on.

If you were fooled

Speed matters more than embarrassment. These scams are built to fool careful people, and a quick report gives banks and investigators a chance.

In the first hour

1. Money sent: Call your bank or payment provider using the number on your card or in your banking app. Ask them to stop or recall the transfer.
2. Password shared: Change it from a device you trust, starting with email. Sign out of all other sessions. Locked Out has step-by-step recovery for major platforms.
3. Code or login prompt approved: Change the password and review recent sign-ins and connected devices.
4. Remote access granted: Disconnect the device from the network. At work, call IT on its known number. At home, have the device checked before using it for banking.
5. At work, report immediately. Tell your security team or manager. A fast report can protect colleagues getting the same call.

Within a day

• Report to law enforcement. In the US, report to the FBI at IC3 with transaction details and a description of the contact, as the FBI asks (FBI IC3). You can also report to the FTC at ReportFraud.ftc.gov.
• Save evidence: numbers, messages, call times and transaction references.
• Warn the impersonated person and your family or team.

Watch for the second scam

Victims are often targeted again by "recovery" services that promise to get the money back for a fee. Legitimate recovery does not work that way. See account recovery scams.

Common mistakes

• Waiting to be sure before calling the bank. Call first and be wrong.
• Staying quiet at work out of embarrassment. A blameless report helps everyone.

Action checklist

• Save your bank's fraud number in your phone now.
• Know who to call at work to report a scam, and save that number.
• If you shared a password, change it from a clean device and sign out all sessions.
• File a report with IC3 or your national fraud agency, with evidence attached.
• Ignore anyone who offers to recover lost money for a fee.

Your complete checklist

• Judge every unexpected request by its ask: money, login, code, file or remote access.
• Agree a family safe word in person and tell every adult in the household.
• Agree a challenge question with your manager and finance contacts.
• Hang up on any urgent call asking for money and call back on a number you already had.
• Never treat a voice or video call as proof of identity on its own.
• Never approve money or credentials on the channel the request arrived on.
• Ask that every payment and payment-detail change at work has a second approver.
• Bookmark login pages for your bank, email and payroll.
• Look up recruiters' companies independently and never pay to get a job.
• Never run interview code or install interview software on a work device.
• Contact support only through the official site or app, and never install remote access because someone called you.
• Never read out a verification code or approve a login you did not start.
• Save your bank's fraud number and your IT help desk's real number today.
• If fooled, call the bank first, change passwords from a clean device, and report to IC3 or your national agency.
• Ignore anyone offering to recover lost money for a fee.