1. Platform
  2. AI-Powered Attacks

AI vishing and deepfakes that think, talk and adapt.

Real attackers now use AI to make phone calls, clone voices, fake video and write emails that answer back. Cyberfrog uses the same techniques, safely and with consent, so your people meet them in a simulation first.

Book a demo See how a call works
  • Live AI phone calls in 28 languages
  • Voice cloning only with consent
  • Deepfake video of your own executives
  • Email and SMS that reply
LIVE AI CALL · IT SERVICE DESK PRETEXTIT Service Desk+41 44 *** 12 0701:42MuteKeypadSpeakerEndTRANSCRIPTAI CALLERHi Anna, Mark from the IT service desk. We just sawa sign-in to your account from Lisbon.EMPLOYEEThat wasn't me. What do you need?AI CALLERI'll block it now. You'll get a code by text,can you read it out to me?EMPLOYEECan I call you back on the service desk number?AI CALLEROf course, that's the right thing to do. Thanks, Anna.LIVE SIGNALSGuardrails onTrustSuspicionGoal progressVerification asked: the AI backs off
dashboard.cyberfrogsecurity.com/attack-templates/new?type=vishingAI voice callSet the goal and the rules. The AI holds the conversation.GOALGet a one-time MFA codePassword-reset codeInstall remote softwareVisit a URLInternal process detailsINFLUENCE TACTICSAuthorityUrgencyFearFamiliarityReciprocityCuriositySocial proofScarcityPERSONAMark Keller, IT service deskKnows the ticket number and Anna's managerAnswers to: who is calling, call me back?LANGUAGEEnglishDeutschFrançais+25GUARDRAILSBack off after 3 refusalsStay in role, never mention AIOff-topic and harmful requests blockedCall ends after 5 minutesTest call to my phoneSave templateScored after every call: disclosed, resisted, and the exact words.
01Live AI phone calls

A phone call that holds a conversation.

An AI caller phones your employee and talks in real time. It listens, answers questions, applies pressure and changes tactics when someone pushes back, the way a social engineer would. When someone does the right thing, it backs off.

Goals
Ten ready goals, from a one-time MFA code or a password-reset code to installing remote software or internal process details. Or describe your own.
Tactics
Authority, urgency, fear, scarcity, social proof, reciprocity, curiosity and familiarity, chosen per scenario.
Languages
28 languages. The caller opens in the employee’s language and can ask which one they prefer.
Voices
Eight natural library voices, or an approved clone of a voice your people know.
  1. Set the goal

    Pick what the caller is after, the persona and the pressure it may use.

  2. The call

    Cyberfrog dials, at a sensible local time, and the AI opens with the pretext.

  3. It adapts

    Questions, pushback and “let me call you back” all get a believable answer.

  4. Scored

    The transcript is judged and the right lesson is assigned.

02Judged by AI, scored by your rules

Every call judged. Word for word.

After the call an AI judge reads the full transcript and marks what the employee gave away and what they protected, quoting their exact words. Your scoring rules set the number, and the follow-up training is assigned automatically.

Score
0 to 100, built from the events you choose and weight: a code read out, an ID confirmed, a callback offered.
Evidence
The full transcript with the moment of each disclosure highlighted. No guessing, no arguing.
Follow-up
A lesson matched to the severity, and an optional note to the manager.
Live signals
During the call Cyberfrog tracks trust, suspicion and progress towards the goal.
dashboard.cyberfrogsecurity.com/campaigns/q3-helpdesk/calls/jonas-weberCall result: Jonas Weber4 min 12 s · English · judged against the goal you setRISK SCORE72/ 100HighTraining assigned:Verify callers (3D lesson)WHAT HAPPENEDDisclosedRead out a one-time code“It's 4 8 2, wait, 9 1 7.”DisclosedConfirmed his employee ID“Yes, that's my number.”ResistedQuestioned the caller once“Why would IT call me about this?”TRANSCRIPTJudged by AI, scored by your rulesAI callerYou'll get a six-digit code by text. Can you read it to me?JonasHang on. It's 4 8 2, wait, 9 1 7.AI callerPerfect, that's done. One more thing: your employee ID?JonasYes, that's my number, 20-4471.
03Voice cloning

A familiar voice, only with consent.

Clone the voice of your CEO, your IT lead or a supplier your people know, so a simulated call sounds like the real thing. Nothing is used before that person has consented and an admin has approved it.

dashboard.cyberfrogsecurity.com/voice-clones/ceoVoice profile: Laura Brandt, CEONo clone is used without recorded consent and approval.1Consent recordedLaura Brandt, laura.brandt@acme.com, ref. HR-2026-1182Consented samples2 recordings, 1 min 40 s3Approved by an admin25 Oct 2026, 10:144ActiveUsable in AI calls and deepfake videosPREVIEW“The board meeting moved to three. Please send me the figures first.”AUDIT LOGvoice.clone.consent_recorded · O. Meiervoice.clone.approved · O. MeierRevoke deletes the voice at the provider
  1. Create a voice profile

    One profile for each person whose voice will be used.

  2. Record consent

    Name, email and a reference, for example the HR ticket that documents the agreement.

  3. Add consented samples

    A few short recordings of the person speaking, in the language of the simulation.

  4. Approve

    An admin approves the profile. Every step is written to the audit log.

  5. Use it

    In AI phone calls and as the voice of a deepfake video.

  6. Revoke at any time

    Revoking a profile deletes the voice at the provider as well.

  • Consent recorded
  • Admin approval
  • Audit-logged
  • Revocable
04Deepfake video

The CEO on video. It isn’t the CEO.

A lip-synced video of a trusted person on a convincing landing page, asking for something urgent. Cyberfrog measures every step: page opened, video watched, button clicked, form submitted.

  • CEO wire transfer
  • IT helpdesk
  • Vendor onboarding
  • HR policy
  • Payroll bank update
  • CEO town hall
  • Security incident
  • Invoice approval
Face
A photo of your executive, uploaded with confirmed consent, a stock avatar or a digital twin.
Voice
A library voice or an approved clone, lip-synced to the script.
Format
16:9, 9:16 or square, 720p or 1080p, with captions, in six languages.
Delivery
A personal, tracked email link for each employee, or one shared page for an awareness day.
board-briefing.acme-group.co/q3-paymentLaura Brandt, CEO · Q3 board update“I need this payment approved before five.”Approve the paymentWatched 50% or more counts as engagedWHAT WE MEASUREPage opened100%Watched half71%Clicked approve38%Submitted the form12%SCENARIO AND FORMATCEO wire transferPayroll bank updateIT helpdeskInvoice approvalTown hall16:99:161:11080pCaptions6 languagesAvatar: your executive's photo with consent, a stock avatar or a digital twin.
05Multi-stage AI conversations

Phishing that answers back.

One opening email, then the AI takes over. It reads every reply, decides whether the employee is engaged, writes the next message and moves to SMS when the story calls for it.

1 OPENING EMAILInterview invitation:Senior Financial AnalystFrom: talent@acme-careers.co2 ANNA REPLIES“Sounds great! What arethe next steps?”Received 09:123 THE AI READS ITClassifiedPositiveEngagedNext stepSend the schedule, then SMS4 AI-WRITTEN FOLLOW-UPThanks Anna! Your interviewschedule is in a secure file.Written for her reply5 ESCALATES TO SMSYour code to open theschedule: 518 204Only aftershe engagedMax 6 AIreplies6 THE FINAL STEPSign in to openthe scheduleCredential page, tracked
  1. 01A playbook, not a free-for-all

    Your playbook sets the stages; the AI writes the messages. Ready-made playbooks: HR interview, IT account check, vendor invoice, payroll portal and secure file share.

  2. 02Every reply understood

    Each answer is classified as positive, negative, neutral or reported, and Cyberfrog records what was shared: a phone number, personal details, a refusal.

  3. 03Knows when to stop

    A cap on AI replies, a timeout, and an immediate stop when someone reports the message. Links and codes are inserted by Cyberfrog, never invented by the AI.

PRETEXTS FOR LENA ROTH, HEAD OF FINANCERisk 81 / 100Left a 5-star review for Ristorante Da GinoSource: Public reviewHigh fitAccount found on a travel booking siteSource: Account enumerationMedium fitEmail address in a 2023 shopping-site breachSource: Breach data, password maskedHigh fitDRAFT EMAILYour table at Da Gino:please confirm the depositConfirm depositPlus a matching web pageWHY THIS WORKS ON LENAFamiliarShe really ate there. The email names the restaurant.PlausibleDeposits for group bookings are normal.Urgent“Held until 6 pm” pushes her to act fast.ReviewableYou approve every draft before it is sent.
06Spear phishing from OSINT

Written from your own footprint.

Point Cyberfrog at a domain. It finds who is exposed, and for the people who matter most an AI writes up to 20 pretexts each, every one tied to a public source. The best become draft emails and web pages that you review.

Sources
Breach data, account enumeration, public reviews and profiles.
Privacy
Passwords are masked before any AI sees them.
Output
A risk score per person, attack vectors ranked by likelihood and impact, and ready drafts.

MSPs can scan prospects and customers on their behalf, and show them what an attacker would see.

More on OSINT campaigns
More AI in every attack

Any attack, in minutes.

The same AI builds and adapts the classic simulations too, so your library keeps up with what attackers send this week.

  • From a prompt

    Describe the attack and the audience, and get an editable phishing email.

  • From a real phishing email

    Upload a screenshot and get a faithful, editable rebuild, optionally translated or rewritten.

  • From this week’s news

    A scam from a public advisory becomes a matching simulation and a short lesson.

  • In 130+ languages

    Translate any email or voice-message template in one step, across a library of 1,000+ attack templates.

  • Matched to each team

    The AI picks the scenario that fits each department and suggests believable senders.

  • A lesson right after

    The simulation itself becomes a short debrief course for everyone who took part.

07Safe by design

Realistic for your people, controlled for you.

AI makes the attacks believable. Your rules decide what it may do, and it stops the moment a simulation would stop being training.

  1. 01Guardrails in every call

    Off-topic and harmful requests are blocked in both directions, and the caller never reveals it is a simulation.

  2. 02Consent before likeness

    No voice or face of a real person is used without recorded consent and an admin’s approval.

  3. 03Checked and logged

    AI requests are moderated and screened for prompt injection, and every use is logged and metered per customer.

  4. 04Your rules decide

    The AI writes and judges. Your goals, playbooks and scoring weights decide the outcome.

Works with the rest of the platform

From the first call to fewer clicks.

Hear an AI call on your own phone.

In a 30-minute demo our AI caller phones you, and we walk through the scored result together.

Book a demo