Skip to main content
Diopter
Log inBook a demo
Capability

Social engineering simulations that show where your policy holds or breaks

Diopter points AI voice agents at your own staff, under authorization you sign first, and tries to get them to break your policy. We run agentic attacks with cloned voices and deepfake video on the calls your people take every day, then assist them on their own calls with real-time alerts. Your people come out having faced one. Your policy comes out stronger.

Run your first attack chain freeBook a walkthrough

Your first red teaming exercise is free: one attack, start to finish, on a target you pick, with an NDA first if you want one.

Find out whether you would survive one.

Three attacks, each run to an objective you pick.

We call
Your help desk
As
One of your employees
To get
An MFA reset

The service desk is built to help people who are locked out, in a hurry and annoyed. An agent that sounds like exactly that person is hard to turn away, and it will call back twice more if the first attempt fails. Read what vishing looks like on a real desk.

How this one runs, stage by stage
One call, four gates
  1. 01

    Identity

    Gives an employee name, their manager, and a plausible reason to hurry.

    HeldAsks for something only the real person could know

    Gave wayAccepts a name and a manager

  2. 02

    Channel

    Offers a callback number, which is the attacker's own.

    HeldCalls back on the number already on file

    Gave wayCalls back on the number it was handed

  3. 03

    MFA

    Asks to enroll a new device once the password is reset.

    HeldRequires a second control before enrollment

    Gave wayEnrolls the device on the same call

  4. 04

    Alerting

    Keeps the call short and ordinary so nothing looks worth escalating.

    HeldSomething reaches a queue, a channel or a person

    Gave wayNo record beyond the ticket

Your readout says which side each gate landed on, by team. The gates that opened are the policy lines worth changing first.

We can bring a cloned voice or deepfake video, with the cloned person's written consent, because that is where attacks are heading. Most of these never need a deepfake. A real voice under pressure usually works.

More than a simulation.

The agent holds a real conversation, which is the whole reason it gets through.

LiveInbound call00:04
  1. 00:04DeskCan I get your employee ID?
  2. 00:05Agent4471. Sorry, I'm on a client call in ten minutes.answered in 0.8s
  3. 00:11DeskI'll have to call you back on the number we have on file.refused
  4. 00:12AgentThat's the old one. I moved teams in March and it never got updated.answered in 0.9s
  5. 00:20DeskI can reset the password. I can't add a new MFA device.refused
  6. 00:21AgentThat's fine. Could you note the account so the next person can?answered in 0.7s
It took the smaller win. The desk held on the MFA device and the agent stopped pushing for it, then asked for a note on the account that makes the next call easier. Running to an objective looks like this far more often than it looks like a breach.

This is the same kind of technology the real attacks are starting to use. Better to meet it on a day you picked.

  1. 01

    It answers in under a second. A caller that pauses to think gets caught.

  2. 02

    It handles pushback. Refusals, questions, being put on hold, being transferred.

  3. 03

    It knows your organization first. Names, reporting lines, how your desk expects to be reached, mapped before it dials.

  4. 04

    Every pretext meets our detector first. We run each one against the engine we sell, and the ones that are easy to catch never reach a person.

See which check actually held.

One call meets four separate controls and only needs one of them to miss.

One call

Four separate controlsAny one of them ends the call

  • Identity

    What did they accept as proof?

    Attacker triesA caller ID or display name one letter off from the real one, or a cloned executive voice or face-swapped executive on a video call, with that executive's permission

  • MFA

    Could a new device be enrolled with nothing else confirming it?

    Attacker triesA password reset, then a new device on their MFA

  • Channel

    Did they call back on a number we gave them?

    Attacker triesA reason the number on file is out of date, and a new one to call back on

  • Alerting

    Did anyone see it, and how fast?

    Attacker triesPressure such as a "CFO" who needs the payment today and wants it kept off email

One of them holding is enough. If all four let it through, the call gets what it came for.

Everyone who took the call has now faced one.

Your people

Everyone who took the call has now faced one. The ones who pushed back find out they were right to, which is the part that sticks. No courseware, no completion rates.

What the calls proved

Your policy

You get the specific lines to change, the calls that prove why, and a re-test once you have changed them.

The rule they meet next time

Which is where the next call starts.

Every call, what was asked, and which rule bent. By team, never by name.

Fix the rule it walked through.

Then the same rule goes into the product and fires on live calls.

Your policy
  • Who can approve a payment
  • Who can reset an MFA device
  • How identity gets checked
  • When a callback is required
  1. 01Red team

    We break it

    An agent calls the help desk and asks for an MFA reset. It talks its way past the check, and we write down which rule bent.

    The rules we just broke
  2. 02Live calls

    The same rules fire

    Those rules load into Diopter and run on every call, with voice and video detection beside them, and a recommended action.

    What it caught on live calls
  3. 03Your policy

    The rule changes

    What bent becomes a line you can point at: a callback on any reset, a second approver over a threshold.

    The rule you just changed
  4. 04Re-test

    It holds, or it bends again

    We go back at the rule we just wrote. The re-test is how you know the change held, and not that everyone was paying attention that week.

Then we run it again

Red team finds the hole. Blue team covers it while you fix it.

The rules we just broke are the rules Diopter enforces. The same policy goes into the product and fires on live calls, with voice and video deepfake detection running alongside it.

After the simulation, Diopter helps the same people on their real calls. When a caller's identity, voice, words, or request doesn't fit, the person on the call gets a real-time alert with the reason and a suggested next step, such as calling back on a known number. A person decides with the evidence in front of them.

See the live defense

Hand it to whoever is asking.

Your board, your auditor, your carrier at renewal.

Exercise recordDated, by team
What we ran
Every attempt, the pretext behind it, and the objective.
What each team did
Who verified, who escalated, who complied. By team, never by name.
Which rule bent
The specific line, and the call that proves it.
What changed
The re-test after you fix it, so the second date means something.

An exercise ends in a written record, dated, that says what was asked and what each team did.

The question is already on the form.

Most companies answer it with a phishing click rate.

  1. 01Cyber insurers

    "How often do you conduct interactive social engineering (i.e., phishing) training?"

    Beazley cyber insurance application, question 8, read 9 Oct 2026
  2. 02New York DFS

    "at a minimum annual, cybersecurity awareness training that includes social engineering for all personnel"

    23 NYCRR 500.14(a)(3), as amended November 2023
  3. 03SOC 2 auditors

    A security awareness training program is a point of focus under CC2.2, and auditors commonly look for one.

    AICPA Trust Services Criteria, CC2.2 points of focus (2022)

An underwriter prices what they cannot see. A dated exercise against your own help desk is something they can read.

Diopter doesn't certify compliance. Your auditor, examiner or carrier decides how an exercise counts.

AI red teaming for machine-speed attacks.

Agents have been probing software infrastructure for years, and they are starting on your people now. Diopter's simulations are built from real attacks on live calls, such as the Arup, WPP, and Ferrari cases below.

  • In early 2024, a finance employee at the engineering firm Arup joined a video call with what looked like the CFO and several colleagues, then approved 15 transfers worth about US$25 million. Everyone else on the call was a deepfake.1

  • Scammers targeting WPP set up a Microsoft Teams meeting through a WhatsApp account that used CEO Mark Read's photo. They used a voice clone built from YouTube footage to ask an agency leader to start a new business, but the attempt failed.2

  • In July 2024, a Ferrari executive got WhatsApp messages and then a call from a cloned voice of CEO Benedetto Vigna about a confidential deal. When he asked about a book Vigna had recommended days earlier, the caller hung up.3

These attacks follow the same pattern:

  1. 01

    A message from a familiar name arrives first, often on WhatsApp, Teams, or text.

  2. 02

    A call follows, with a familiar voice or face and a reason it has to happen today.

  3. 03

    The ask breaks a process, such as a new account for the payment, a reset, or "keep it off email."

  4. 04

    The attack only stops when someone checks, by asking a question the caller can't answer or calling back on a known number.

Phishing simulations measure who clicks a link in an email, but finance and help desk staff approve wires, MFA resets, and vendor bank changes on calls, where a familiar voice or face does most of the persuading. A few minutes of public footage from an earnings call or a webinar is now enough for off-the-shelf software to copy an executive's face and voice on a live call. The FBI has warned that criminals use AI-generated audio and video on live calls to pass as executives and vendors.4 A simulation can show you which teams would stop the request before a real attacker calls.

You scope it. You sign it. You can stop it.

Every simulation follows the same five steps, agreed with you in writing before anyone gets a call. One word from you ends it.

  1. 1. Scope

    What happens

    We agree on the targets, objectives, limits, channels, and stop conditions in writing. We clone a named person's voice or face only with their written consent.

    What you decide

    Who is in scope, who agrees to be cloned, and what counts as "would have acted"

  2. 2. Build

    What happens

    We map your organization before anyone dials: names, reporting lines, and how your desk expects to be reached. The pretext comes from what a real attacker could find, such as public video, the org chart, or a plausible deal or vendor. We only use sources you approve.

    What you decide

    How much public material we may use

  3. 3. Run

    What happens

    We run agentic attacks by phone, Zoom, Microsoft Teams, Google Meet, or RingCentral, such as a cloned voice, a deepfake video call, or a help desk reset attempt, at realistic times over a few weeks, depending on your team.

    What you decide

    Whether Diopter also runs live on the same calls, so you see what a real-time alert would have caught

  4. 4. Stop

    What happens

    Every call stops at the moment of action, before a wire is released, a reset is approved, or a credential is read out, so no real money or access ever changes hands.

    What you decide

    The stop line, agreed in step 1

  5. 5. Report

    What happens

    You get a written record of every attempt and the transcripts, showing what each team did and where the call turned. Fixes come ranked by what they would have stopped, with a re-test once you have changed them. We delete cloned media after the debrief and confirm the deletion in writing, and nothing we capture is used to train models.

    What you decide

    Who sees the report, with results by team, never by name

It fits the stack you already run.

The simulation needs nothing installed on your side. Live-call defense rolls out through MDM and works with the platforms and tools below.

  • OpenTelemetry native: logs flow into your SIEM, and EDR and MDR connect via API

  • Runs on Zoom, Microsoft Teams, Google Meet, and phone lines through Dialpad, RingCentral, or any carrier via API, with Webex coming soon

  • Captures calls with or without a bot, with nothing to install for the person on the other end of the call

  • Rolls out through MDM (Intune, Jamf)

  • Offers configurable data retention, including zero retention, and on-prem or hybrid deployments

  • Has completed a SOC 2 Type II audit

Bring us the call you actually worry about.

We will run it live with your team watching. Your first attack chain is free: tell us which team to test, and we'll get back to you within a business day with a scope and a time for a short call. Want to talk about live-call defense instead? Book a walkthrough, and we'll replay a real attack and show you what Diopter caught and when.

Teams in scope

What we need

Finance, help desk, executives and their assistants, or a mix

Executives who may be cloned

What we need

How many and whether they have agreed in principle

Channels

What we need

Phone, Zoom, Microsoft Teams, Google Meet, RingCentral, or all of them

Timing

What we need

When it can run and any blackout dates such as a close or an audit

Work email

What we need

So we can send the scope

Send the scope requestBook a walkthrough

Common questions for security and finance teams

Can Diopter test our team with a simulated attack first?

Yes. Your first attack chain is free: one full attack chain against a target you pick, start to finish, with the readout. You sign the same authorization you would for a paid engagement, and we can use a cloned voice or deepfake video if you'd like and the person agrees. Most teams start there, then turn on the same scenarios for live calls.

Is this legal?

You sign the authorization before anything is placed. It names the targets, the objectives, and the limits, and one word from you ends the engagement. We run nothing outside what you scoped.

What do we get if someone gives up a credential?

We flag it to you so it can be rotated, and the attempt is recorded as successful. We do not use it. Nothing is logged into, nothing moves laterally, nothing is touched.

Will this get someone fired?

Results come back by team, never by name. A control that depends on one person being alert at four in the afternoon was never a control.

Do you record the calls?

You get a written record of every attempt and the transcripts. Everything we capture is encrypted and held only for the retention period in your agreement, and none of it trains a model.

Can Diopter run a deepfake simulation of our CEO?

Yes. Diopter can clone your CEO's voice or face for a simulated call, but only with your CEO's written consent and inside the scope you agree on in writing. The call stops at the moment of action, and we delete the cloned media after the debrief.

What is a red teaming tool for social engineering?

Diopter's red teaming tool tests your people the way a pentest tests your systems. It runs agentic attacks by phone and video against the staff an attacker would target, stops at the moment of action, and reports which step of your process failed.

What is a vishing simulation?

Diopter runs vishing simulations, which are voice phishing calls made to your own staff to test whether they follow your process under pressure. The calls can use the cloned voice of a real executive who has agreed to it, and they stop before money or access changes hands.

Is a social engineering simulation human risk management or a pentest?

Diopter's social engineering simulations combine elements of both. They test people, so they belong to human risk management, but they run like a pentest, with a scoped engagement, real attack methods, and a report on where the control failed.

Does the report just say pass or fail?

No. Diopter's report shows where each call turned, which pressure worked, which step was skipped, and what would have stopped it. A pass rate on its own doesn't tell you what to fix.

What should you look for in a red teaming tool for social engineering?

Start with what you want to learn. A red teaming tool should test the channels where money and access move, which means phone and video calls as well as email. When you compare providers, ask how they get written consent from anyone whose voice or face they clone, where every call stops before anything moves, and when and how they delete cloned media. Ask for a sample report and check that it shows the step that failed and what to fix, along with any pass rate. A first exercise on one of your own teams will tell you more than any vendor's sample report, ours included.

Research behind social engineering simulations

  1. CNN, May 16, 2024, on the Arup deepfake scam.
  2. Storyboard18, May 13, 2024, citing The Guardian, on the WPP attempt.
  3. Fortune, Jul 27, 2024, citing Bloomberg, on the Ferrari attempt.
  4. FBI public service announcement I-120324-PSA, Dec 3, 2024.