Deepfake-assisted phishing
TL;DR: Voice cloning needs 30 seconds of audio; video deepfakes need a minute or two of source footage. In 2024-25 real engagements began incorporating: (1) voice-cloned CFO calling AP for wire transfers, (2) live deepfake video on Zoom for “executive direction” SE, (3) AI-generated phishing copy at scale, (4) AI-generated supporting evidence (fake LinkedIn profiles, screenshots). Defence shifts from “spot the grammar errors” to “verify out-of-band”. Companion to phishing-infrastructure-design and pretext-design-for-engagements.
What’s now feasible
| Capability | What you need | Output quality |
|---|---|---|
| Voice clone | 30s of clean audio | indistinguishable on phone in casual context |
| Real-time voice over phone | clone + Voicebox / ElevenLabs / RVC | near-real-time with 200-500ms latency |
| Live video deepfake on Zoom | source video + DeepFaceLive / FaceFusion | passable in low-light; tells visible on closer look |
| Style-matched phishing copy | LLM + a sample of target’s tone | indistinguishable from human |
| Synthetic employee LinkedIn | LLM + synthesised profile picture | passes casual scrutiny |
Real-world incidents (public)
- 2019: UK CEO scam — voice-cloned executive directed €220k wire transfer.
- 2024: Arup HK engineer joined a “Zoom call” with multiple deepfaked executives, transferred $25M.
- Many smaller incidents at SMB scale; under-reported.
Attack chain — voice-clone wire transfer
- Recon. Identify the target — typically AP or treasury staff who can authorise wires.
- Voice harvest. Pull 30-60s of the impersonated executive’s voice from:
- YouTube earnings calls, conference talks.
- Company podcast / town hall.
- Voicemail (in some cases, public).
- Clone. ElevenLabs / RVC / Tortoise-TTS produces a voice model.
- Timing. Off-cycle (Friday afternoon, executive away on travel).
- Pretext. “Urgent — I’m with a vendor at a contract closing — wire $X to $Y by 5pm, will email confirmation”.
- Spoofed CallerID. VOIP services let you spoof; arrives as “CEO_NAME on CallerID”.
- Execute. AP processes wire.
- Cover. “Don’t email me; I’m in meetings” — kills the verification step.
Defence
- Out-of-band verification. Pre-agreed phrase / code for wire approval, communicated via separate channel.
- Multi-party authorisation. No single approver for large wires.
- Time-delay. Wires above threshold delayed 4-24 hours.
- Caller ID training. Staff trained that Caller ID is not authentication.
- Voice biometric — limited availability; can be evaded but adds friction.
Attack chain — live Zoom deepfake
- Source video. YouTube interview, internal video shared externally.
- Train. DeepFaceLive / SimSwap / FaceFusion. A consumer GPU produces real-time face-swap.
- Hardware setup. Webcam → OBS Virtual Camera (with face swap applied) → Zoom.
- Voice clone for matching audio.
- Call victim. Schedule the meeting via a plausible-looking calendar invite.
- Direct. “Confirm this transaction now; don’t loop anyone else in.”
Tells (still detectable in many cases):
- Slow blink rate, unusual lighting on face boundary.
- Reflections in glasses don’t match.
- Hand gestures partially occluding face cause artefacts.
- Voice latency mismatch.
Defence:
- Pre-scheduled meetings only from trusted calendar systems.
- Multi-party for high-stakes decisions.
- Liveness challenges — “Touch your left ear” (deepfakes struggle with novel hand-on-face).
- Skepticism training; “did this person just ask me to authorise a wire?” → escalate.
Attack chain — AI-generated phishing at scale
LLM-written phishing:
- Tailored per target (LinkedIn-mined details).
- Grammar perfect.
- Matches the target company’s tone.
- Generated at the rate of one per second.
A phishing campaign that used to take a week of copywriting can be generated in an hour.
Defence:
- Email gateways now use LLM-vs-LLM detection (less effective; arms race).
- Banner labels on external mail still help.
- Phishing-simulation training that includes AI-generated lures.
Attack chain — synthetic employee
Job-search platforms used by attackers:
- Apply for a remote role with a synthetic profile.
- Pass video interview using deepfake.
- Get hired; insider threat from day one.
Reports from 2023-25 indicate this is happening at meaningful frequency, particularly for software roles.
Defence:
- Live in-person onboarding before access provisioning.
- Background checks beyond ID verification.
- Behavioural-anomaly detection on early-tenure employees.
Attack chain — synthetic evidence for OSINT poisoning
Attacker generates supporting evidence for a scam:
- Fake LinkedIn profiles of co-workers / vendors.
- Synthesised GitHub commit history.
- Fake reference letters.
- AI-generated company website at a typo-squat domain.
When the victim “verifies” by Googling, the search results look consistent.
Defence:
- Don’t accept LinkedIn as authoritative.
- Use known-good directories (vendor records, customer-listed contacts).
- Multi-source verification.
Engagement use (pentest)
If your engagement letter explicitly authorises social engineering with synthetic media:
- Document the source video clip used.
- Capture audio waveform of cloned voice for evidence.
- Don’t conduct real transfers; demonstrate to the point of “would have happened”.
- Don’t impersonate a specific executive without explicit written permission from that executive.
Many engagement letters now have a “no deepfake of executives by name” clause.
Detection — defender side
- Liveness checks in video calls (Zoom is adding native; vendor-specific).
- Voice biometric matching (limited deployment).
- Synthetic-media detection APIs (Reality Defender, Sensity, Microsoft Azure AI Content Safety).
- OOB verification policy.
OSCP/OSEP/OSWE relevance
Out of scope. Important for SE engagements and real-world risk modelling.
Legal
Voice-cloning a specific person without their consent for fraudulent purposes is illegal in most jurisdictions. Engagement-letter authorisation must be explicit and specific. Cross-border (USA SAG-AFTRA, EU AI Act, deepfake laws in China, South Korea) varies widely.
References
- FBI — Voice and video deepfake advisories
- Microsoft — Trust no AI / deepfake research
- Verizon DBIR — annual SE and BEC data
- DeepFaceLive — research only
- ElevenLabs — voice cloning (commercial, with safety review)
- See also: phishing-infrastructure-design, pretext-design-for-engagements, ai-agent-confusion-attacks, multimodal-attacks, client-side-attacks-primer