AI-Generated Phishing and Deepfake Vishing: What We’re Actually Seeing in Live Incidents

Every vendor’s Cybersecurity Awareness Month content this year has a version of the same slide: “AI is supercharging phishing.” It’s not wrong. It’s also become background noise, a scary number pulled from a press release, dropped into a blog post, repeated by the next vendor down the line until nobody can trace it back to an actual study. Before adding to that pile, it’s worth separating what’s rigorously documented from what’s marketing, and grounding both in what CyberClan’s incident responders are actually being called in to clean up. 

What the Data Actually Supports 

Start with the claim that gets repeated most: that the vast majority of phishing is now AI-generated. A widely circulated figure that 82.6% of phishing emails contain AI-generated elements, has no transparent methodology behind it and should be treated as directional at best, not a verified statistic. We’d rather tell you that than repeat it. 

Here’s what does hold up. A controlled study by Heiding et al. found that AI-generated phishing emails achieved a 54% click-through rate, statistically matching campaigns written by human social-engineering experts, against a roughly 12% baseline for generic, non-personalized phishing. The same research put the production cost at approximately $0.04 per email using freely available open-source intelligence to personalize each message. That’s the real story: AI hasn’t necessarily made phishing more effective than a skilled human attacker, it’s made human-expert-level personalization available at a cost of essentially nothing, at scale. 

The vector shift is where 2026 data gets more interesting than the email story alone. Mandiant’s M-Trends 2026 report found email phishing accounted for only 6% of initial infection vectors, a significant decline, while voice phishing rose to 11%, becoming the second most common initial-access method behind vulnerability exploitation (32%). Attackers are moving up the trust ladder: a cloned voice on a phone call carries more built-in credibility than a written email, and it’s now cheap enough to industrialize. 

Deepfake Vishing Is Moving from Novelty to Standard Toolkit 

The volume numbers here are genuinely striking. CrowdStrike’s 2025 Global Threat Report documented a 442% surge in vishing activity between the first and second half of 2024, and by H1 2025, voice-based attack volume had already exceeded the full-year 2024 total. Pindrop’s fraud research found deepfake fraud attempts in contact centers rose more than 1,300% in 2024, from roughly one incident a month to seven a day, with the contact-center industry now facing an estimated $44.5 billion in fraud exposure for 2025, and AI-generated content driving 42.5% of fraud attempts overall. 

Two publicly documented incidents illustrate what this looks like at the high and low end of success: 

The one that worked: In January 2024, engineering firm Arup lost $25.6 million (HK$200 million) when attackers combined a spear-phishing email impersonating the CFO with a live video conference featuring multiple deepfaked executives. The attackers built the fakes from publicly available source material, LinkedIn videos, conference recordings, virtual meeting footage, and used the video call’s apparent internal consensus to authorize 15 fraudulent wire transfers in a single day. The fraud wasn’t caught by any technical control; it surfaced only when the employee who’d approved the transfers mentioned the “secret transaction” to headquarters during a routine follow-up. 

The one that didn’t: In May 2024, scammers targeted WPP using an AI voice clone of CEO Mark Read, built from public YouTube footage, combined with a spoofed Microsoft Teams call and a fake WhatsApp account to make the approach look coordinated and legitimate. The attempt to extract money and personal information from other WPP executives failed, the target organization’s staff didn’t take the bait. 

The gap between those two outcomes usually isn’t better deepfake-detection technology. Detection remains genuinely hard: research on automated deepfake detection shows accuracy dropping by 45–50% when moving from lab conditions to real-world calls, and humans identify deepfakes correctly only 55–60% of the time, barely better than a coin flip. The gap is process: whether someone had a reason, and the authority, to stop and verify through a second channel before money moved. 

What This Looks Like Inside a Real Engagement 

The following is a composite pattern drawn from recurring elements across multiple CyberClan incident response engagements, not a single client’s incident, to illustrate the pattern without disclosing any identifiable details. 

A mid-market company’s finance team receives a call that appears to come from a known vendor’s accounts-receivable contact, using a voice that sounds correct because it was cloned from a public webinar recording the real contact had given months earlier. The caller references a real, recent invoice, pulled from an earlier, more ordinary phishing email that had already compromised a mailbox, and asks the finance employee to update banking details “before the next payment run” due to a supposed bank transition. There’s urgency, but not the kind that trips alarm bells: no threats, no strange requests, just a plausible-sounding administrative update from someone who sounds exactly like who they claim to be. The update goes through. The next legitimate payment to that vendor goes to the attacker instead. 

Nothing in that pattern requires a video call or a sophisticated real-time deepfake — a cloned voice reading a script over a phone line was enough, because the actual point of failure was a payment-detail change processed without an independent, out-of-band verification step. That’s the pattern worth internalizing: most successful AI-enabled social engineering CyberClan sees isn’t defeating a technical control, it’s exploiting the absence of a verification step that would have taken thirty seconds. 

What Actually Reduces This Risk 

Awareness training helps people recognize a suspicious email. It’s less effective against a confident voice on a phone call referencing real, accurate details. The controls that actually move the needle on deepfake vishing and AI-generated phishing look different: 

• Out-of-band verification for anything involving money or credentials. Any request to change payment details, wire funds, or reset access should require confirmation through a separate, previously known channel, not a callback number provided in the same call or email making the request. 

• Verbal or code-word verification for executive requests. A pre-agreed verification phrase, changed periodically, is a low-tech control that neutralizes most voice-cloning attempts because the attacker has no way to know it in advance. 

• Segregation of duties on financial approvals. The Arup case succeeded partly because one employee could be convinced to authorize 15 transfers in a day. Requiring independent second approval for high-value or unusual transfers closes that gap regardless of how convincing the request sounds. 

• Treat helpdesk and vendor calls as a distinct attack surface. Attackers increasingly target IT helpdesks and password-reset workflows with vished social engineering. Helpdesk staff need explicit verification procedures for identity resets, not just general phishing awareness. 

• Update your incident response and tabletop scenarios to include voice and video, not just email. Most tabletop exercises still script a phishing-email scenario. If your last exercise didn’t include a deepfake voice or video call, it’s testing last decade’s threat model. 

Talk to Someone Who Sees This Firsthand 

AI phishing 2026 and deepfake vishing aren’t hypothetical trends for Cybersecurity Awareness Month content calendars, they’re active techniques in ransomware and fraud cases CyberClan’s incident response team is working right now. If you want to pressure-test your organization’s verification procedures against these patterns before an attacker does, schedule a Cyber Resilience Assessment or download our Vishing & Deepfake Response Playbook. 

Think you’re on a call with a deepfake right now, or already seeing suspicious wire activity? Call our 24/7 emergency hotline immediately, every minute matters once funds are moving.