Deepfake Phishing Simulation Software: How to Train Teams

What is deepfake phishing simulation software?

Deepfake phishing simulation software is a training technology that safely recreates AI-generated voice and video attacks, so employees can practice spotting them without any real risk. Employees receive a phishing email that leads to a fake Teams, Meet, or Zoom call, where an avatar with a cloned voice and face urges an action; if the employee takes the bait, that’s a safe fail: It immediately triggers a short, focused micro-training that builds resilience against these advanced attacks.

Platforms like Hoxhunt support multi-channel environments and can deliver either targeted benchmarks or broad awareness modules.

Key capabilities you should expect

• Multi-step, multi-channel flow: An email leads to a live call simulation, mirroring real attack paths.
• Realistic avatars: Each one uses a cloned voice and face, built only with consent, from providers we’ve checked to make sure they never store or train on your data.
• Instant micro-training: Employees get in-context coaching the moment they take a risky action, with no public shaming, just fast feedback.
• Responsible realism: The simulation can add light glitches or lag to make a “bad connection” pretext feel real, without ever crossing psychological-safety lines.

How it differs from traditional phishing simulations

• It covers voice and video tricks, not just text or email.
• It trains people to question a suspicious request rather than trying to spot a fake video or voice: deepfakes can already look and sound real, but an unusual request stays unusual no matter how convincing the fake is.
• It focuses on high-value moments, like payments, access requests, and approvals, where a deepfake attack pays off most for attackers.

‍

Inside Hoxhunt’s deepfake simulation workflow

Hoxhunt’s deepfake phishing simulation mirrors real attacks: A phishing email routes the user to a fake Teams, Meet, or Zoom page, a cloned-voice avatar urges an urgent action, and clicking the link in the chat triggers a safe fail with instant micro-training.

Step-by-step flow (multi-channel simulation)

1. Phishing email (entry point): Users receive a realistic phishing simulation email (Outlook or Gmail) prompting a quick call. Hoxhunt’s reporting button can capture reports directly from the inbox.
2. Attacker-controlled “meeting” (landing page): Clicking opens a browser page that convincingly mimics Microsoft Teams, Google Meet, or Zoom; camera and mic can be enabled, which gives it a realistic touch. Any links on the page look legitimate, mirroring what a real attacker-controlled page would show.
3. AI-generated voice and avatar (deepfake voice and video): The avatar urges an urgent action, for example: opening a “SharePoint” link.
4. User action A → safe fail: If the user clicks the link in the chat, it triggers a pop-up saying it was a deepfake fraud simulation, followed by an instant micro-training. This is how Hoxhunt’s security awareness training coaches away risky behavior: no public shaming, just fast, contextual feedback.
5. User action B → threat reporting path: If the user suspects it was an AI-generated voice phishing, they can report the original email by clicking the Hoxhunt button. It still counts as a successful report if made before clicking the link in the chat, even if the user has clicked the link in the email (the one that opens the deepfake call).
6. Admin visibility & outcomes: Admins can view campaign metrics in real time and trigger behavior-based nudges or training packages for users who show risky behavior.

Watch our live demo of Hoxhunt’s deepfake simulation.

KPIs to track (beyond “click rate”)

Use these three KPIs to quantify behavior change and reduce human cyber risk in multi-channel deepfake phishing simulation exercises.

Why the realism works but stays responsible

• Web environments closely mirror Teams, Meet, or Zoom (familiar communication channels), including camera and mic prompts and domain styling.
• The experience stays real but responsible. It’s brief and targeted. It always ends in micro-training that builds resilience against deepfake scams, focusing on changing the user’s behavior instead of making them feel guilty about having made a mistake.

Below you can walk through our deepfake simulation training at your own pace and see how it works under the hood.

Realism without risk: consent, ethics, and safety

Security teams need realistic simulations, not a set of “gotcha” tricks. That’s why Hoxhunt’s deepfake training only runs with the target’s consent, follows a pre-written script, and stays brief. It plays out in an attacker-style fake meeting that ends with micro-training designed to change behavior. It never makes the employee feel bad about a mistake. Reporting happens from the original email, and rollouts are targeted one-offs with spacing, followed by additional training for risky users.

1) Consent & legal guardrails (non-negotiable)

• Contractual cover: Customers sign a deepfake amendment, unless the main contract already includes clauses covering deepfakes.
• Customer-obtained consent: The customer confirms that the person being cloned has consented (one-line confirmation is sufficient).
• Providers & data use: We use ElevenLabs for voice and Synthesia or HeyGen for video, and our agreements specify that providers may not store customer content or train models on it.

2) “Real but responsible” employee experience

• Cloning always requires consent, and scenarios are pre-scripted. They never turn into open-ended chats.
• Interactions stay short and respectful. They never turn into a drawn-out back-and-forth.
• Clicking the link in the chat leads to an “Oops, that could have been dangerous” learning moment. This way, employees get a quick piece of micro-training that, instead of punishing them, helps them learn from their mistake in a fun way.

3) Designed limits that keep risk low

• Browser-based environment: Simulations run in browser-based look-alikes of Teams, Meet, or Zoom; we can add light lag or glitch effects to sell the pretext, but we cap the realism there, so the experience never feels more stressful than necessary.
• Clear end state: Clicking the link in the chat is the defined “fail,” which triggers instant coaching. Exiting the call before clicking avoids the fail and counts as a good outcome.
• Reporting path: Employees report from the original email using the Hoxhunt button; there’s no way to report from inside the fake meeting page. A report still counts as success if made before clicking the link in the chat.

4) Cadence and content strategy (avoid training plateaus)

• Rollouts run as a targeted one-off simulation for a single group, such as finance, executive assistants, or approvers, then let the buzz settle before re-testing during regular training refreshers.
• Additional training ties directly to user behavior, such as fails or risky clicks. It uses short, ready-made modules instead of long-form courses, keeping things lightweight.

‍

Use cases & rollout: where deepfakes fit in your phishing campaigns

Deepfake simulations work best where real-world threats hurt most: executive impersonation and business email compromise (BEC) approvals, vendor fraud (an attacker posing as a supplier to redirect a payment), and IT support pretexts. Success means better reporting rate, faster drop-off, and fewer fails across your phishing campaigns and simulation programs.

High-impact use cases (pair with realistic scenarios)

• Executive impersonation and BEC approvals: A cloned CEO or CFO on a quick call pushes an urgent wire or bank-detail change, a classic business email compromise pattern. Then, in the chat, they drop a link that leads to a credential harvester.
• Vendor fraud in finance: Attackers send a fake “updated remit info” or “invoice about to bounce” message by email, then follow up with a call, targeting approvers and accounts payable (AP) analysts.
• IT support scams: Attackers send a “Reset now, verify device” message inside a fake Teams, Meet, or Zoom room, and then drop a SharePoint or SSO (Single Sign-On) look-alike link in the chat.
• Multi-channel pretexting: Sequencing an email with a follow-up call boosts credibility, especially when voice messages are added to mirror AI-powered voice phishing.

Who to target first (and why)

• Finance, executive assistants, and approvers face the highest monetary risk from fraud attempts and BEC.
• IT and service desk face credential theft and policy bypass pressure.
• The broader organization faces lower risk, but still gets short training modules to build baseline recognition.

Rollout pattern we recommend

• Targeted benchmark (one-off): Launch it to a defined cohort.
• Space it: Let the buzz cool before delivering micro-training or ready-made modules to those who struggled.
• Re-test: Run a new scenario through the same process, and look for improvement over time in measurable behavior: more reports, more safe drop-offs, and fewer fails.
• Close the loop: Use analytics to trigger additional training, keep ongoing training lightweight, and avoid long-form defaults that cause training plateaus.

‍

Getting started: deployment, consent & customization

A deepfake phishing simulation program can go live in just weeks, with a small team and a short setup process. Here’s how it comes together.

Step-by-step rollout

1. Choose the scenario & cohort: Start where business email compromise or approvals hurt most (finance and AP, EAs, approvers; IT for credential resets). Keep the scenario anchored to real-world threats.
2. Consent & contract: Confirm that the person’s likeness and voice are used with their consent. Bear in mind that Hoxhunt’s providers do not train models or store your content.
3. Capture assets: Record 1–2 minutes of the subject’s voice (or use approved samples), then supply a still image or video plus a script. There are two avatar options: fully consented capture (the most realistic) or a photo plus a short voice sample (the lower-effort route).
4. Production window: The standard turnaround is about 5 business days after we receive the voice sample, plus about 1 day for avatar generation.
5. Launch the drill (realistic but safe): Send a simulated phishing email that routes to a browser-based Teams, Meet, or Zoom look-alike, where the cloned-voice avatar triggers the same safe-fail flow described earlier.
6. Define success before go-live: Remember the key metrics from earlier, defined here for this specific rollout:
   • Reporting rate (from inbox): users who report via the original email before clicking the link in the chat.
   • Drop-off before action: users who exit the fake call without clicking.
   • Final fail rate: users who click the link in the chat (the endpoint).
7. Close the loop with training: Trigger micro-training or training modules for users who showed risky behavior during the drill.
8. Re-test for proof: Space out simulations, swap in a fresh scenario, and re-run it to prove fewer phishing clicks and real behavior change across your campaigns.

‍

Deepfake phishing simulation vendor comparison table

A snapshot of where Hoxhunt sits against other deepfake and phishing-simulation vendors, based on public product information.

‍

How attackers use deepfakes in phishing today

Attackers combine voice cloning and avatar video with multi-channel pretexting to impersonate trusted people. They scrape org intel (like LinkedIn or email signatures), spoof caller ID or domains, lure targets into a fake Teams, Meet, or Zoom page, then push urgent actions such as credential theft, a fraudulent wire transfer, or vendor fraud.

Encoder-decoder architecture

The process starts with the encoder, a neural network that compresses input data, a face image, for example, into a smaller representation called a latent space. This latent space keeps only the most essential features (shape, structure, texture, expression, angle, lighting) and drops the rest, which is what lets the model efficiently manipulate and recombine those features to generate new outputs.

The decoder then takes that compressed representation and reconstructs it into a full image; for deepfakes, it takes the latent representation of one face and generates the face of another, effectively swapping features between the two.

Generative adversarial networks (GANs)

Deepfake technology relies on a type of deep learning called Generative Adversarial Networks (GANs), built from two neural networks:

During training, the Generator and the Discriminator are pitted against each other: The Generator tries to create increasingly convincing fakes while the Discriminator gets better at spotting them. Over time, that back-and-forth sharpens the Generator until it produces images or videos that the Discriminator, and increasingly the human eye, can no longer tell apart from the real thing.

The AI is trained using datasets

The GAN is trained on vast datasets of real images, audio, or video, depending on the type of deepfake being created. To build a deepfake video of a specific person, for example, the model trains on thousands of images or clips of that person’s face, capturing a range of angles, expressions, and lighting conditions. The more diverse and comprehensive that training data is, the more convincing the result: The GAN keeps iterating on the deepfakes it produces until the Generator’s output is virtually indistinguishable from the real thing.

And these datasets are then used to generate deepfakes

Once trained, the GAN can generate deepfake content: manipulating a video by altering the subject’s facial expressions, syncing their lips to different audio, or replacing their face entirely with someone else’s. The result looks authentic, even though it has been artificially created or manipulated. GANs are typically used in a couple of different ways:

What signals they exploit (why people fall for it)

• A familiar voice adds authority and urgency (e.g., “I’m boarding, do this now”).
• Familiar communication channels, like Microsoft Teams, Google Meet, or Zoom, lower suspicion.
• People still widely trust caller ID and domain names today.
• Remote and hybrid workflows push quick approvals off the email trail, where they’re harder to verify.

Deepfake statistics

The voice-cloning survey figures below were reported across 2024 and 2025 and remain the most-cited public benchmarks as of 2026; treat them as a general trend rather than fresh data.

• 70% of people say they aren’t confident that they can tell the difference between a real and cloned voice.
• 53% of people share their voices online or via recorded notes at least once a week.
• Searches for “free voice cloning software” rose 120% over the past year.
• Three seconds of audio is sometimes all that’s needed to produce an 85% voice match.
• 43% of business email compromise attacks now include a callback phishing lure that draws the victim into a fraudulent follow-up phone call (Hoxhunt Phishing Trends Report 2026).

These trends are accelerating: Hoxhunt’s 2026 data found AI-generated phishing surged roughly 14× at the end of 2025, jumping from under 5% to 56% of detected attacks in a single month (Hoxhunt Phishing Trends Report 2026). That means the realism captured in the statistics above is no longer an edge case; it is becoming the default an employee will face.

The capability gap is closing on the attacker side too: In Hoxhunt testing, AI spear-phishing agents now beat elite human red teams (professional simulated-attack specialists) at getting people to click (Hoxhunt Phishing Trends Report 2026, p. 49). When artificial intelligence can craft a more convincing lure than a trained professional, employees need to practice against that same level of realism.

What this means for defenses & training

That gap needs to close on the training side too. Here’s what to do about it:

• Train reflexes and ‘the pause’ over visual scrutiny: Assume visuals can look perfect.
• Rehearse multi-channel scenarios inside your security awareness program (email → meeting → request).
• Measure what matters in deepfake phishing simulations: reporting rate (from inbox), drop-off before action, and final fail rate, and use behavior-based nudges for users who exhibit risky behavior.

‍

Proof that realistic, high-volume practice works

What really changes employee behavior? It’s repeated practice against realistic lures. That kind of varied, repeated exposure is what prepares a workforce for deepfake-grade attacks.

What to teach employees about deepfake phishing

The best approach is to coach employees to verify the premise. Spotting flaws in the video isn’t enough anymore, because deepfake quality keeps improving, and a video or voice clone that looks flawless is no longer a reliable signal on its own.

1) Teach the reflexes (deepfakes can get around typical red flags)

• Warning signs: The message uses authority and urgency, makes an out-of-process ask (like payments or credential verification), or arrives through an unusual channel for the task.
• Coach the reflex: Ask “Why me, why now, why this channel?”, then verify out-of-band, through a different channel than the one the request came in on.
• Rationale: As deepfake quality rises, glitch-spotting becomes unreliable, but an odd, out-of-process request is still a red flag no matter how convincing the video looks.

2) Teach voice/call protocols (AI-generated voice phishing)

• Caller ID: Treat it as untrusted; spoofing internal names or extensions still works today.
• Verification playbook: Hang up, call back via the directory, or message on Slack or Teams; agree on passphrases with frequent contacts in advance.

3) Teach meeting/link hygiene (fake meeting rooms)

• Pattern to expect: An email lure leads to a browser-based Teams, Meet, or Zoom clone that asks for camera and microphone access the same way a real video call would, which is what makes the fake meeting convincing.
• Golden rule: The link in the chat is always the payload, often a SharePoint or SSO look-alike page. Exiting before clicking counts as success.

4) Teach the reporting flow (maps to KPIs)

• Report from inbox: In Hoxhunt, users can report via the Hoxhunt button on the original email; it still counts as success even if they already joined the call, which increases the reporting rate.
• Safe exits: Leaving the fake call without clicking is a good outcome that increases drop-off before action.
• Avoid the click: Do not open the link in the chat, which lowers the final fail rate.

5) Teach process boundaries for high-risk workflows

• Finance & vendor fraud: Never approve an urgent wire transfer or bank-detail change on a live call; re-route to the documented process instead.
• IT support scams: Never share credentials over voice; use official reset flows instead.
• Executive impersonation: Verify unexpected deepfake CEO messages out-of-band.

‍

Why deepfake phishing needs specialized simulation tools

Most phishing simulation tools teach link-spotting in email. Deepfake attacks weaponize artificial intelligence across channels: email lure → fake Teams, Meet, or Zoom call → urgent request from a cloned voice or avatar. To build strong cybersecurity habits, programs must rehearse these realistic scenarios, measure behavior (report, drop-off, fail), and refresh training content beyond generic templates.

What basic programs miss

• Email-only focus: Basic programs stay email-only, but real attackers chain channels, so simulations need to mirror that multi-vector realism (inbox → meeting clone → link in chat).
• Visual-glitch bias: Basic programs lean on “spot the glitch” training, but as deepfake quality rises, that stops working. Checking the premise of a request still holds up, even under pressure.
• Training plateaus: Basic programs rely on stale, generic templates and long-form modules, which stall engagement and learning. A rotating library of templates and short, focused lessons keeps it fresh instead.

‍

To build the broader security awareness program around these defenses, read our phishing simulation best practices playbook, compare platforms in our guide to the best phishing simulation tools, and see how the same AI powers AI phishing attacks across the inbox.

Sources

Deepfake phishing simulation software FAQ