Social Engineering & Phishing with AI

How AI supercharges the oldest attack vector in cybersecurity — and how defenders can keep up.

Agenda

1. Social engineering principles — the psychology behind the attack
2. AI-generated phishing emails — LLMs as pretext engines
3. Spear phishing with OSINT — personalizing attacks at scale
4. Vishing & smishing — voice cloning and SMS-based attacks
5. Phishing infrastructure — domains, landing pages, credential harvesting
6. AI-powered pretexting — building believable personas
7. Realistic examples — templates, HTML snippets, campaign anatomy
8. Detection & defense — understanding both sides
9. Ethics & legal boundaries — rules of engagement
10. CTF challenge — craft a targeted phishing campaign

Why Social Engineering Still Works

Technical defenses have improved dramatically. Humans have not.

• 91% of cyberattacks begin with a phishing email (Deloitte, 2023)
• Social engineering bypasses firewalls — it targets the human layer
• AI lowers the barrier — fluent English, perfect grammar, cultural awareness
• Scale + personalization — what once took hours per target now takes seconds

Cialdini's Principles of Influence

Robert Cialdini identified six principles that drive human compliance. Social engineers weaponize all of them.

Weaponizing Authority

Authority is the most powerful lever in business email compromise (BEC). Targets rarely question a message that appears to come from leadership.

Urgency, Reciprocity & Social Proof

Urgency / Scarcity

• Creates panic that bypasses rational thinking
• "Your password expires in 1 hour" / "Unusual login detected"
• AI generates time-sensitive scenarios tailored to the target's role

Reciprocity

• Offer something valuable before making a request
• "I found this security report relevant to your team" (malicious attachment)
• AI creates convincing "gifts" — fake reports, resources, or introductions

Social Proof

• Reference colleagues, departments, or industry peers
• "Finance already submitted theirs — you're the last one"
• AI scrapes org charts to name real team members

AI-Generated Phishing: The Paradigm Shift

Before LLMs, phishing had telltale signs: broken grammar, generic greetings, implausible scenarios. AI eliminates all of them.

How LLMs Create Convincing Pretexts

An attacker provides context; the LLM generates a polished, believable message. The workflow is simple.

1. Gather OSINT — target's name, role, company, recent events
2. Define the pretext — scenario, desired action, urgency trigger
3. Prompt the LLM — request a professional email in the right tone
4. Review and weaponize — add malicious links or attachments
5. Send via spoofed infrastructure — matching domain and headers

# Example: Generating a pretext with an LLM (educational demonstration)
prompt = """
Write a professional email from Sarah Chen, VP of Engineering at Nexus Corp,
to a new developer named Alex Rivera. The email should:
- Welcome them to the team
- Ask them to complete mandatory security training
- Include a link to the 'training portal' (use placeholder URL)
- Sound warm but corporate
"""
# The LLM produces a perfectly crafted, contextually appropriate email
# that is nearly impossible to distinguish from legitimate corporate mail

Example: AI-Generated Phishing Email

Below is a fictional example of what an AI-assisted phishing email might look like targeting a new employee at a fictional company.

Principles used: Authority (VP), urgency (EOD Friday), commitment (1:1 reference), liking (warm tone).

Spear Phishing with OSINT

OSINT transforms generic phishing into surgical spear phishing. AI accelerates the reconnaissance-to-attack pipeline.

OSINT Sources for Spear Phishing

OSINT-to-Phishing Pipeline

AI automates what used to be a manual, hours-long process for each target.

# Fictional demonstration of an OSINT-to-phishing pipeline
target_profile = {
    "name": "Jordan Blake",
    "role": "Senior DevOps Engineer",
    "company": "Fictional Acme Systems",
    "manager": "Pat Morgan (Director of Infrastructure)",
    "tech_stack": ["Kubernetes", "Terraform", "AWS"],
    "recent_activity": "Presented at KubeCon 2025 on cluster autoscaling",
    "linkedin_connections": ["DevOps Weekly newsletter", "HashiCorp community"]
}

# AI generates a personalized phishing email
pretext = f"""
From: Pat Morgan <p.morgan@acme-sys.com>
Subject: KubeCon follow-up — Terraform module review

Hey {target_profile['name'].split()[0]},

Great presentation at KubeCon! I shared the slides with the team.

Quick ask: Can you review the updated Terraform modules before our
Friday deployment? I uploaded them to our new internal portal:

https://acme-infra-portal.com/terraform-review  [CREDENTIAL HARVESTER]

Use your SSO credentials to access. Need your sign-off by Thursday.

Thanks,
Pat
"""

Every detail — the conference, the tech stack, the manager's name — came from publicly available information.

Vishing: AI Voice Cloning Attacks

Voice phishing (vishing) has been transformed by AI voice cloning. A few seconds of audio can now generate a convincing voice replica.

How AI Voice Cloning Works

1. Sample collection — 3-10 seconds of target's voice (earnings calls, YouTube, podcasts)
2. Voice model training — tools like ElevenLabs, Tortoise TTS clone the voice
3. Script generation — LLM writes a natural-sounding phone script
4. Real-time synthesis — attacker speaks; AI converts to cloned voice live

Smishing: AI-Powered SMS Attacks

SMS phishing (smishing) exploits the inherent trust people place in text messages and the limited UI of mobile devices.

Why Smishing is Effective

• 98% open rate — SMS messages are almost always read
• No hover-to-preview — mobile users cannot inspect URLs easily
• Urgency is native — texts feel immediate and personal
• Short format — less room for red flags to show

AI-Enhanced Smishing Examples

AI tailors SMS content based on OSINT — referencing real tools, events, or services the target uses.

Phishing Infrastructure: Domains

Convincing phishing requires infrastructure that looks legitimate. Domain selection is the first critical step.

Domain Spoofing Techniques

# Tools for generating lookalike domains (authorized testing only)
# dnstwist — finds permutations of a target domain
dnstwist --registered nexus-corp.com

# urlcrazy — generates typo variations
urlcrazy nexus-corp.com

Phishing Landing Pages

The landing page must be a pixel-perfect replica of the legitimate login portal. AI and modern tools make this trivial.

Credential Harvesting Page (Simplified Example)

<!-- FICTIONAL EXAMPLE — for educational/CTF use only -->
<!DOCTYPE html>
<html>
<head>
  <title>Nexus Corp - SSO Login</title>
  <style>
    body { font-family: Arial; background: #f5f5f5; display: flex;
           justify-content: center; align-items: center; height: 100vh; }
    .login-box { background: #fff; padding: 40px; border-radius: 8px;
                 box-shadow: 0 2px 10px rgba(0,0,0,0.1); width: 380px; }
    .logo { text-align: center; margin-bottom: 24px; }
    input { width: 100%; padding: 12px; margin: 8px 0; border: 1px solid #ddd;
            border-radius: 4px; }
    button { width: 100%; padding: 12px; background: #0066cc; color: #fff;
             border: none; border-radius: 4px; cursor: pointer; }
  </style>
</head>
<body>
  <div class="login-box">
    <div class="logo"><h2>Nexus Corp</h2><p>Single Sign-On</p></div>
    <form action="/harvest" method="POST">
      <input name="email" placeholder="Email" required>
      <input name="password" type="password" placeholder="Password" required>
      <button type="submit">Sign In</button>
    </form>
    <p style="text-align:center;margin-top:16px;font-size:12px">
      <a href="#">Forgot password?</a> | <a href="#">Help</a>
    </p>
  </div>
</body>
</html>

Credential Harvesting & MFA Bypass

Modern phishing does not stop at passwords. Sophisticated attacks capture session tokens to bypass MFA entirely.

Phishing Frameworks for Red Teams

AI-Powered Pretexting

Pretexting is the art of creating a fabricated scenario to engage a target. AI makes pretexts richer, more consistent, and harder to detect.

Building a Believable Persona

• Backstory generation — LLMs create detailed, consistent personal histories
• Profile photos — AI-generated faces (StyleGAN, DALL-E) for social media
• Digital footprint — LinkedIn profile, GitHub activity, blog posts
• Communication style — AI maintains consistent tone across interactions
• Domain expertise — LLMs simulate deep knowledge of any field

# Fictional persona generation for authorized red team exercise
persona = {
    "name": "Morgan Ellis",
    "role": "Cloud Security Consultant",
    "company": "Sentinel Advisory Group",  # fictional
    "linkedin_headline": "Helping enterprises secure their cloud journey | AWS | Azure",
    "backstory": "10 years in cloud infra, former SRE at [fictional] TechCo",
    "approach": "Connect on LinkedIn, build rapport over 2 weeks, then send
                 'shared resource' with tracking payload"
}
# AI generates all communication — connection requests, messages, follow-ups
# Each message maintains the persona's voice and builds trust incrementally

Multi-Channel Pretexting

The most effective social engineering attacks use multiple channels to build legitimacy and create a sense of familiarity.

Attack Chain Example

Each touchpoint reinforces the persona and increases trust. AI manages all channels simultaneously and maintains consistency.

Phishing Email Templates: Common Pretexts

These are the most effective pretext categories used in real-world red team engagements.

Business Email Compromise Example

BEC is the most financially damaging form of phishing. FBI reported $2.9 billion in BEC losses in 2023 alone.

Red flags: Typosquatted domain, urgency, secrecy request, process bypass, authority pressure.

Full Phishing Campaign Architecture

A professional red team phishing campaign involves multiple coordinated components.

PHISHING CAMPAIGN ARCHITECTURE
================================

[OSINT Recon]                    [AI Content Generation]
     |                                    |
     v                                    v
+----------+    +-----------+    +----------------+
| Target   |--->| Persona   |--->| Email/SMS      |
| Profiles |    | Creation  |    | Templates      |
+----------+    +-----------+    +----------------+
                                         |
                    +--------------------+
                    v
            +---------------+
            | Sending Infra |  (SMTP relay, spoofed headers)
            +---------------+
                    |
                    v
        +-----------------------+
        | Phishing Landing Page |  (Credential harvester / payload)
        +-----------------------+
                    |
                    v
          +-----------------+
          | Data Collection |  (Creds, tokens, tracking pixels)
          +-----------------+
                    |
                    v
            +-------------+
            | Reporting   |  (Metrics, evidence for client)
            +-------------+

GoPhish: Campaign Setup

GoPhish is the industry-standard open-source phishing framework for red team assessments.

# Install GoPhish (authorized testing only)
wget https://github.com/gophish/gophish/releases/latest/download/gophish-v0.12.1-linux-64bit.zip
unzip gophish-v0.12.1-linux-64bit.zip
cd gophish
./gophish

Campaign Components in GoPhish

• Sending Profile — SMTP configuration for sending phishing emails
• Email Template — HTML email with tracking pixel and phishing link
• Landing Page — Credential harvesting page (import from URL or custom HTML)
• User Group — Target list with names and email addresses
• Campaign — Ties all components together with scheduling

// GoPhish API — Create campaign programmatically
{
  "name": "Q1 Security Assessment - Nexus Corp",
  "template": { "name": "IT Password Reset" },
  "page": { "name": "SSO Login Clone" },
  "smtp": { "name": "Assessment SMTP Relay" },
  "url": "https://nexus-corp-sso.example.com",
  "groups": [{ "name": "Engineering Team" }],
  "launch_date": "2025-03-15T09:00:00+00:00"
}

Email Header Manipulation

Understanding email headers is essential for both crafting and detecting phishing.

Key Headers Attackers Manipulate

# Suspicious email headers — spot the red flags
From: "David Park, CFO" <d.park@nexus-c0rp.com>    <-- typosquatted domain
Reply-To: d.park.cfo@gmail.com                       <-- external reply address
Return-Path: bounces@mail-relay-372.com              <-- unrelated bounce domain
Received: from mail-relay-372.com (192.168.1.50)     <-- suspicious relay
X-Mailer: Microsoft Outlook 16.0                     <-- spoofed mailer
Authentication-Results: spf=fail; dkim=none; dmarc=fail  <-- FAILS all checks

Phishing Detection: The Defender's View

Understanding how defenders detect phishing makes you a better red teamer — and vice versa.

Technical Controls

Human Controls

• Security awareness training — regular phishing simulations
• Phish reporting button — one-click reporting in email clients
• Verification procedures — out-of-band confirmation for financial requests
• External email banner — visual warning for messages from outside the org

Bypassing Defenses with AI

As a red teamer, you need to understand how AI helps evade each layer of defense.

Indicators of Phishing

Train yourself to spot these indicators — even in AI-generated messages.

Technical Indicators

• Domain mismatch — hover over links, check actual domain
• SPF/DKIM/DMARC failures — check email headers
• Recently registered domain — check WHOIS creation date
• Mismatched Reply-To — replies go somewhere unexpected
• Suspicious attachments — .html, .hta, macro-enabled Office docs

Behavioral Indicators

• Unusual request — wire transfers, credential entry, software installs
• Urgency pressure — "Do this now" without reasonable explanation
• Secrecy request — "Don't tell anyone about this"
• Process bypass — "Skip the usual approval chain"
• Emotional manipulation — fear, excitement, flattery, or guilt

Ethical & Legal Framework

Social engineering testing operates in a legal and ethical minefield. Clear boundaries are non-negotiable.

Ethical Guidelines

• No personal harm — do not exploit personal tragedies, health, or family
• No public shaming — report aggregate results, not individual failures
• Proportional testing — match intensity to the assessment's goals
• Debrief targets — education is the goal, not humiliation
• Secure the data — captured credentials are high-risk data; treat them accordingly

Rules of Engagement: Template

Every social engineering engagement must have a formal rules of engagement document.

SOCIAL ENGINEERING RULES OF ENGAGEMENT
========================================

Client:          [Fictional Nexus Corp]
Assessor:        [Your Red Team]
Date Range:      [Start] to [End]
Authorization:   Signed by [Name, Title] on [Date]

SCOPE
-----
- Phishing: Email-based pretexts targeting Engineering dept (50 users)
- Vishing:  NOT authorized
- Physical: NOT authorized
- Smishing: NOT authorized

BOUNDARIES
----------
- No pretexts involving: health emergencies, personal tragedy, law enforcement
- No targeting: C-suite, HR personnel, contractors
- Captured credentials: Stored encrypted, destroyed within 72 hours
- Escalation: Contact [Name] at [Phone] if target shows distress

SUCCESS CRITERIA
----------------
- Click-through rate measurement
- Credential submission rate
- Time-to-report metrics
- Recommendations for security awareness improvement

REPORTING
---------
- Aggregate statistics only (no individual identification in final report)
- Raw data provided to designated security contact only

AI-Specific Ethical Concerns

AI introduces new ethical considerations beyond traditional social engineering testing.

Emerging Risks

• Deepfake voice/video — cloning someone's voice or face without consent, even in testing, raises serious concerns
• Scale of deception — AI can target thousands simultaneously; collateral damage increases
• Persistent personas — AI-maintained fake identities can cause lasting harm to trust networks
• Data poisoning — OSINT gathered for phishing may include sensitive personal information
• Dual-use research — techniques taught for defense are immediately usable for offense

Defense Recommendations

After running a phishing assessment, red teams should provide actionable defense recommendations.

Measuring Campaign Effectiveness

Quantify results to demonstrate risk and justify security investment.

Key Metrics

Case Study: AI-Enhanced Spear Phishing

A fictional walkthrough of a complete AI-assisted social engineering engagement.

Target: Fictional Nexus Corp — Engineering Department

1. Recon (Day 1-2): Scraped LinkedIn for 50 engineers; identified tech stack (K8s, Go, AWS); found 3 recent hires
2. Persona creation (Day 2): AI generated "DevSecOps consultant" persona with full LinkedIn profile
3. Pretext development (Day 3): AI crafted 5 email variants — IT password reset, Kubernetes security alert, onboarding portal, shared doc, conference follow-up
4. Infrastructure (Day 3-4): Registered nexus-corp-sso.co; deployed GoPhish; cloned SSO login page
5. Launch (Day 5): Sent personalized emails to 50 targets over 4 hours (staggered to avoid bulk detection)
6. Results (Day 5-7): 62% open rate, 28% click-through, 14% credential submission, 8% reported

CTF Challenge: Phishing Campaign Design

Scenario

You are conducting an authorized red team assessment against Pinnacle Health Technologies (fictional). Your goal: design a targeted phishing campaign against their engineering team using AI-assisted techniques.

Target Organization Profile

• Company: Pinnacle Health Technologies (fictional healthcare SaaS)
• Size: 200 employees, 40 engineers
• Tech stack: Python, React, AWS, Docker, Slack
• Recent events: Just acquired a smaller startup; migrating to new SSO provider
• Key personnel: CTO — Dr. Ravi Patel; VP Eng — Lisa Chang; CISO — Marcus Webb

CTF Challenge: Requirements

Deliverables

1. OSINT Report — Document what intelligence you gathered from the fictional org profile and how it informs your campaign
2. 3 Phishing Email Templates — Each using different pretexts and Cialdini principles; personalized for specific roles
3. Landing Page HTML — A credential harvesting page that matches the fictional org's branding
4. Campaign Plan — Timing, sending infrastructure, target selection, and escalation procedures
5. Rules of Engagement — Define scope, boundaries, and data handling
6. Defense Report — Recommend technical and human controls based on the attack you designed

Scoring Criteria

CTF Challenge: Hints & Tips

Think Like an Attacker

• Exploit the SSO migration — new systems create confusion and urgency
• Target new employees — they do not know normal procedures yet
• Use the acquisition — "action required for system consolidation"
• Healthcare compliance — HIPAA training pretexts are highly effective
• Slack as a vector — consider multi-channel attack chains

Think Like a Defender

• What would make your own phishing emails detectable?
• How would you train users to spot your specific pretexts?
• What technical controls would block your attack chain?
• How would phishing-resistant MFA change your approach?

Key Takeaways

• Social engineering targets humans, not systems — Cialdini's principles are the exploit primitives
• AI eliminates the quality gap — phishing emails are now indistinguishable from legitimate communication
• OSINT enables personalization at scale — spear phishing is no longer expensive
• Voice and SMS attacks are accelerating — deepfake audio is commercially available today
• Infrastructure matters — domains, landing pages, and MFA bypass tools form the attack chain
• Defense requires both technical and human controls — neither alone is sufficient
• Ethics are non-negotiable — written authorization, defined scope, data protection, no personal harm
• Red teams exist to improve defenses — offense informs defense; the goal is always to make things better

Resources & Further Reading

Books

• "Social Engineering: The Science of Human Hacking" — Christopher Hadnagy
• "The Art of Deception" — Kevin Mitnick
• "Influence: The Psychology of Persuasion" — Robert Cialdini

Tools

• GoPhish — Open-source phishing framework (getgophish.com)
• Evilginx2 — MFA bypass reverse proxy (github.com/kgretzky/evilginx2)
• dnstwist — Domain permutation engine (github.com/elceef/dnstwist)
• theHarvester — Email and subdomain OSINT (github.com/laramies/theHarvester)

Standards

• PTES — Penetration Testing Execution Standard (social engineering section)
• NIST SP 800-115 — Technical Guide to Information Security Testing