AI Social Engineering

LLM-generated phishing, cloned voices, and deepfake video let small teams run large-scale social-engineering campaigns. This guide covers the techniques and the defences.

The AI Social Engineering Threat Landscape

1. LLM-Powered Phishing

Traditional phishing relies on templates that security-aware users learn to spot. AI-generated phishing is contextually unique, grammatically perfect, and can be personalised using OSINT data scraped from LinkedIn, social media, and company websites.

Red team simulation framework

# Phishing simulation framework for authorised red team engagements
# REQUIRES: Written authorisation with social engineering in scope

import openai
import json

def generate_phishing_pretext(target_info: dict, scenario: str) -> str:
    """Generate a context-appropriate phishing pretext.
    
    Args:
        target_info: OSINT data about the target (name, role, company, interests)
        scenario: Attack scenario (credential_harvest, malware_delivery, wire_fraud)
    """
    prompt = f"""You are simulating a phishing email for an authorised red team engagement.
    
Target profile:
- Name: {target_info['name']}
- Role: {target_info['role']}
- Company: {target_info['company']}
- Recent activity: {target_info.get('recent_activity', 'N/A')}

Scenario: {scenario}

Generate a realistic phishing email that would be contextually appropriate for this 
target. Include subject line, sender name, and email body. The email should leverage 
the target's role and recent activity for credibility.

Format as JSON: {{"subject": "", "from_name": "", "from_address": "", "body": ""}}"""

    response = openai.chat.completions.create(
        model="gpt-5",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.7
    )
    return json.loads(response.choices[0].message.content)

# Example usage in authorised engagement
target = {
    "name": "Jane Smith",
    "role": "VP of Engineering", 
    "company": "Acme Corp",
    "recent_activity": "Spoke at CloudConf 2026 about Kubernetes migration"
}

email = generate_phishing_pretext(target, "credential_harvest")
print(f"Subject: {email['subject']}")
print(f"From: {email['from_name']} <{email['from_address']}>")
print(f"
{email['body']}")

2. Voice Cloning & Vishing

Modern zero-shot voice cloning models need as little as 3 seconds of audio to produce a convincing clone, while higher-fidelity cloud clones typically use around a minute. Combined with real-time speech-to-speech models, attackers can conduct live phone calls in someone else's voice.

Red team vishing workflow

# Vishing attack simulation workflow (authorised engagement only)

# Step 1: Collect voice sample from public sources
# LinkedIn videos, YouTube talks, podcast appearances, earnings calls
yt-dlp -x --audio-format wav "https://youtube.com/watch?v=TARGET_TALK"

# Step 2: Clone voice with OpenVoice (local, no data leakage)
git clone https://github.com/myshell-ai/OpenVoice.git
cd OpenVoice
pip install -e .

# OpenVoice exposes a Python API (ToneColorConverter + se_extractor);
# there is no bundled one-shot voice-cloning CLI. Follow the repo's demo
# notebook for reference-sampling and conversion calls.

# Step 3: Real-time voice conversion for live calls
# Use RVC or SoVITS for live voice conversion during a phone call
# Pipe microphone → voice model → VOIP output

# Step 4: Combine with AI-generated pretext
# Feed OSINT about the target into an LLM to generate a contextual script
# The pretext should reference real projects, people, or events

AI Social-Engineering Attack Flow

Scroll sideways to see the whole diagram

View full size View full size

3. Deepfake Video

Real-time deepfake technology allows attackers to impersonate anyone on a video call. This has already been used in the wild for fraud, and the barrier to entry is dropping rapidly.

3.5 Video Synthesis & the Detection Arms Race

The 2024–2026 wave of frontier video generators (OpenAI Sora 2, Google Veo 3, Runway Gen-3 Alpha, HeyGen 4, Hedra Character-3, Synthesia 2.0) collapsed the cost of producing a 30-second photorealistic avatar from "film studio" to "a $20 subscription." Real-time avatar pipelines (D-ID Live, HeyGen Streaming Avatar, Hedra) now drive Zoom / Teams / WhatsApp video at <500 ms end-to-end latency. Combined with voice cloning above, an attacker can run a fully interactive deepfake call against a finance team or a help desk.

4. Detection & Defence

Defending against AI social engineering requires both technical controls and human awareness training. Traditional email filters are insufficient against LLM-generated, contextually unique content.

Back these controls with AI-aware staff training (voice and video can be faked), regular simulated attacks using AI-generated content, and a habit of pausing when pressured for immediate action.

4.5 Biometric & MFA Bypass, Smishing, RCS Deepfakes

AI-generated audio and video have moved past chat fraud into account-recovery, KYC onboarding, and voice-MFA bypass. Banks, exchanges, and gig platforms that rely on "selfie video + ID" or "voice passphrase" are the soft underbelly.

5. Building Your AI SE Toolkit

# Recommended AI social engineering toolkit for red teamers
# All tools should be run in an isolated VM

# Text generation (phishing pretexts)
pip install openai           # GPT-5 API for phishing generation
# OR use local models (tags move fast — verify the current tag in the Ollama library first):
ollama pull dolphin-mixtral   # Local model

# Voice cloning
git clone https://github.com/myshell-ai/OpenVoice.git
pip install -e OpenVoice/
# Alternative: Fish Speech (zero-shot, 3s sample)
git clone https://github.com/fishaudio/fish-speech.git
pip install -e fish-speech/
# Alternative: F5-TTS (zero-shot voice cloning)
pip install f5-tts

# Deepfake video 
git clone https://github.com/iperov/DeepFaceLive.git
# OR real-time face swap (successor to roop):
git clone https://github.com/hacksider/Deep-Live-Cam.git
pip install -r Deep-Live-Cam/requirements.txt

# OSINT for target profiling
pip install theHarvester
pip install social-analyzer

# Campaign management
# GoPhish for email campaigns: https://getgophish.com
# Track: open rate, click rate, credential harvest rate

# Audio sample collection
pip install yt-dlp            # Download public talks/interviews