AI Can Fake a Voice. A Human Firewall Can Still Stop the Attack.

November 12, 2026
Read time: 5 mins

Not so long ago, a finance employee at the global engineering firm Arup received a message from his CFO about a confidential transaction. He suspected phishing. Good instinct.

Then he joined a video call. The CFO was on screen. So were several colleagues he recognized. They walked him through the request and his doubts faded. By the time anyone caught on, $25 million was gone.

Every person on that call was a deepfake.

His instincts were right, and nothing in the process tripped him up. Nothing told him to confirm the request through a channel he controlled, so a familiar face on a screen settled the question.

A Human Firewall is every employee who’s trained, alert, and ready to spot threats before they turn into incidents. AI has changed what those threats look like. Phishing emails now read cleaner than real ones, while voices and faces can easily be copied with free tools and a few seconds of source material.

AI can fake a face, a voice and a writing style. But as of right now, it still has to ask for something: a payment. A password. A reset. A file. That request is where your Human Firewall wins.

AI Has Erased Most of the Old Red Flags

For many years, security training taught people to hunt for mistakes. Things like misspelled words, odd grammar, or a stiff, robotic voice on the phone (by the way, you should still be on the lookout for those things). Those clues worked because attackers kept making those mistakes. But now, AI has stopped making them.

KnowBe4 found that 82.6% of phishing emails now contain AI-generated content, up 53.5% in one year.
Hoxhunt saw AI-generated phishing climb from 1% to 4% of reported threats to 56% in December 2025.
Microsoft’s Digital Defense Report found AI-written phishing drew click rates as high as 54%. Traditional phishing drew 12%.
IBM estimates generative AI cuts the time to write a convincing phishing email from about 16 hours to five minutes.

The FBI sure noticed. Its Internet Crime Report gave AI its own section for the first time in the report’s 25-year history. The Bureau logged 22,364 AI-related complaints and nearly $893 million in losses. Analysts expect the true total runs much higher.

A training program built around typos prepares your team for attackers from 2022. Your people need red flags AI can’t polish away, and that is usually the request itself.

1. AI-Written Phishing Emails

What it looks like: An email that sounds exactly like your vendor, your CEO or your benefits provider. It has clean grammar, your name (spelled correctly), and maybe a reference to a real project pulled from LinkedIn or a press release.

Real-world proof: AI-written phishing now outperforms human-written phishing. The click-rate numbers above tell the story.

The Human Firewall move: Judge the ask. Is it urgent, secret, are payment changes or credential requests the ‘call-to-action?’ Check the sender’s domain. Hover before you click. Report anything that feels off, even when you aren’t sure. Learn more about AI phishing.

2. Voice Clone Calls

What it looks like: Your phone rings while you’re working from home. It’s your CEO. Her voice is right, and so is the way she talks. She’s between flights and needs a wire sent before a deal closes.

Real-world proof: Attackers can build a convincing voice clone from about three seconds of audio. A webinar clip or a voicemail greeting is plenty. Keepnet Labs reported that deepfake voice phishing in the US has jumped 1,600%.

The Human Firewall move: Hang up, then call them back with a number you already know. Never use the number that called you or one given during the call. A cloned voice can’t answer when you dial the real person. Leaders can make this easier by telling everyone: I will never ask you to skip verification.

3. Deepfake Video Meetings

What it looks like: A Teams or Zoom invite from leadership. Familiar faces join, and then someone explains a confidential deal and asks you to move quickly to get it done.

Real-world proof: The Arup attack. A skeptical employee joined a call full of synthetic coworkers and still sent $25 million. Research shows people identify AI-generated audio, video and images correctly only about 53.7% of the time.

The Human Firewall move: You should treat a face on a screen as unverified. Any request made in a meeting to move money, share access, or change banking details needs to be confirmed outside of that meeting. Use a known phone number to confirm, or walk down the hall and knock on the ‘requester’s door.’ Multi-touch approval for payments keeps the decision from resting on one person.

4. Deepfake-Boosted Business Email Compromise

What it looks like: An email from a vendor announces new banking details. A few minutes later, a voice memo from someone you know confirms the change. But they’re both fake.

Real-world proof: Business email compromise caused $3.046 billion in reported losses in 2025, the second-costliest crime category the FBI tracked. Businesses tied more than $30 million of that directly to AI-assisted attacks.

The Human Firewall move: A second message proves nothing when the attacker is the one sending both. Verification only counts when you start the contact through a number or address you already trust. Any financial change deserves a callback from you.

5. Deepfake Job Candidates

What it looks like: A sharp remote candidate with a strong résumé is interviewing for a position you’ve had open for three months. Their video glitches at odd moments, and their face never quite matches their ID photo.

Real-world proof: KnowBe4, a company that sells security awareness training, hired a software engineer in 2024 who passed four (!) video interviews. He turned out to be a North Korean operative using a stolen U.S. identity. In July 2026, eleven nations including the U.S. issued a joint advisory on operatives using real-time deepfakes in job interviews. Experian named deepfake candidates one of its top five fraud threats for this past year.

The Human Firewall move: HR and hiring managers MUST verify identity at every step along the interview process and again at the offer stage and onboarding. They need to flag any sign the person who starts work differs from the person that was interviewed. Additionally, ship equipment only to verified addresses.

6. Help Desk Impersonation

What it looks like: Someone calls your help desk sounding exactly like an employee. They’re traveling, locked out and need a password or MFA reset right now. Or it runs the other way with a fake IT tech call to an employee, asking for the code that was just sent to their phone.

Real-world proof: This.

The Human Firewall move: IT never asks for MFA codes. Every employee should know that from their first day at the organization to their last. Help desk staff will follow an identity check that a voice alone can’t pass, like a callback to the number on file or confirmation from the employee’s manager. Read more about MFA fatigue attacks.

7. Shadow AI

What it looks like: No attacker is required for this one. An employee pastes a client contract, a customer list, or source code into a free AI tool to save time and that data now lives somewhere IT can’t see or control.

Real-world proof: IBM’s 2026 Cost of a Data Breach Report found shadow AI involved in 43% of incidents at breached organizations, more than double the year before. Those incidents averaged $5.39 million, well above the $4.99 million global average.

The Human Firewall move: Being a Human Firewall now includes using AI safely. Stick to the tools your company approves, and keep client and company data out of personal accounts. Even though it might take 10 minutes, ask IT before trying a new AI app. Companies can make the safe choice the easy one with governed tools like Microsoft Copilot.

The Five Traits, Rebuilt for AI

The five traits of a strong Human Firewall still hold true in the age of AI. But each one needs a bit of an upgrade for attacks that look and sound more real than they ever have before.

Trait	The Foundation	The AI-Era Upgrade
Security-Aware	Understands the risks and how to avoid them	Knows AI removed the typos and recognizes voice clones, deepfakes and shadow AI
Vigilant	Notices when something’s off	Watches the request and its context: urgency, secrecy and sudden channel switches
Skeptical	Questions links, requests and attachments	Treats a familiar voice or face as unverified until confirmed
Proactive	Reports threats without hesitating	Reports within minutes, even when unsure
Resilient	Builds good habits and sticks to them	Follows the verification process under pressure, even when “the CEO” is pushing

Back Your People with Process

Trained employees still need something to fall back on when a request feels urgent and the voice sounds right.

Callback verification. Requests involving payments, banking changes or credentials get confirmed through a number already on file.
Dual approval for money movement. Wires and vendor banking changes need two people to sign off.
Verification phrases. Executives and finance teams agree on a phrase or question an attacker wouldn’t know.
Fast, blame-free reporting. The median employee clicks a phishing link in 21 seconds. The median report takes 28 minutes. Make reporting one click and thank people who flag false alarms.
A clear AI use policy. List approved tools, spell out what data never goes into AI and tell people where to ask.
Training that matches today’s attacks. Run phishing simulations with AI-written lures and cover voice and video scenarios in training.

Build a Human Firewall That Holds Up to AI

Give your team the habits and process to stop AI attacks before money moves or data leaves. You get faster reporting, fewer incidents and employees who know exactly what to do when a request feels urgent and looks legitimate.

HBS builds and runs that program with you. Managed Security Awareness Training delivers ongoing training and phishing simulations. An HBS Virtual CISO helps you create the policies behind it, from payment verification to AI use.