Key takeaways:
- AI is supercharging old social engineering tactics - attackers use AI-powered voice calls to impersonate help desks and get past MFA codes and credentials.
- The threat is growing fast, and causes serious damage. Voice phishing is now the second most common initial access vector.
- Process controls – like out-of-band verification, ID checks and phishing-resistant authentication – hold up even when the call is convincing enough to foll a well-trained employee.
Share this post:
Subscribe:
Get the latest news and insights from Integreon delivered to your inbox.
With AI, what’s old can be new again.
While it’s apparent that AI is capable of incredibly sophisticated attacks these days, threat actors are finding just as much success using it to make old-school social engineering techniques even more effective.
One of the most pervasive and successful tactics right now is IT help desk impersonation, where attackers use AI-powered calls to convince employees to share MFA codes or other credentials. What’s particularly concerning is how convincing these interactions have become. AI can adapt the conversation in real time, create a sense of urgency, and say exactly the right thing to disarm the person on the other end.
For example, an AI caller might say, “I know this isn’t our normal protocol, but the threat actor is in your inbox right now, and every second counts. You’re welcome to hang up and call IT directly if you’d prefer.”
It’s that combination of urgency, apparent transparency, and seemingly reasonable suggestions that make these attacks so effective. The underlying technique isn’t new, but AI is making it far more sophisticated, scalable, and difficult to recognize.
The Data Confirms What We're Seeing in the Field
This isn’t an isolated pattern. Mandiant’s M-Trends 2026 report found that voice phishing is now the second most common initial access vector overall, appearing in roughly 11% of investigated intrusions, trailing only vulnerability exploitation. In cloud compromises specifically, voice phishing (also known as “vishing”) was the single most common way in, ahead of every other vector.
The growth curve is just as telling. CrowdStrike tracked a 442% surge in vishing in the second half of 2024, followed by another 134 percent year-over-year increase into 2025, and its most recent threat hunting data shows call volumes doubling again in the first half of 2026. Separately, deepfake-enabled vishing attempts surged more than 1,600 percent in a single quarter, and roughly one in four people tested could not reliably distinguish a cloned voice from a real one.
Once a caller gets past the help desk, the damage compounds fast. In investigations tied to the threat cluster behind several of the highest-profile retail and hospitality breaches of the past two years, Mandiant found attackers moving from initial contact to full domain administrator access in approximately 40 minutes.
Why the Technique Still Works
Help desk impersonation isn’t succeeding because the technology fooling people is new. It’s succeeding because it exploits something structural: help desks exist to be helpful, and the entire job is built around removing friction for employees who are locked out or under pressure. AI didn’t invent that dynamic. It just removed every tell that used to expose it: the hesitation in an unfamiliar voice, the accent that didn’t match the employee’s file, the script that broke down under follow-up questions.
Several of the highest-profile breaches in the past two years, including major incidents at MGM Resorts, Caesars Entertainment, and a wave of UK retailers, trace back to this exact pattern: a convincing call to the help desk, a credential or MFA reset granted in good faith, and a threat actor inside the environment before anyone realized what had happened.
What This Means for How Organizations Prepare
The instinct is to treat this as a training problem, and to teach employees to recognize suspicious calls. That instinct isn’t wrong, but it’s incomplete. When the caller sounds exactly like a real colleague or a real IT technician, and knows enough real details to sound credible, individual vigilance stops being a reliable control.
The more durable fix sits at the process level: out-of-band verification for any credential or MFA reset tied to a high-value account, identity verification that doesn’t rely on information an attacker could plausibly already have, and phishing-resistant authentication methods that can’t be talked around on a phone call no matter how convincing the caller sounds. Organizations that have built these checkpoints into their help desk workflows are measurably harder targets, not because their employees are better trained, but because the attack has nowhere left to work.
The Broader Pattern
Help desk impersonation is one example of a wider shift worth naming directly: AI is not primarily creating new categories of attack. It’s taking techniques that already worked and is removing the friction that used to limit how often they succeeded. That distinction matters how organizations prioritize defense. The question isn’t only “what new attack is AI capable of.” It’s “which of our existing, human-dependent controls were we relying on friction to protect, and what happens now that the friction is gone.”
That’s the lens worth applying across every category of social engineering risk, not just the help desk, as AI continues to close the gap between a scripted con and a genuinely convincing one.
