How AI Changed the Social Engineering Playbook
Older scams often gave themselves away. Misspellings, awkward phrasing, and robotic voices were once reliable warning signs.
Generative AI has removed many of those tells. According to the 2026 DBIR, the median threat actor researched or used AI assistance in 15 documented attack techniques, with some applying it to as many as 40 or 50.
The bigger change isn’t any single capability. AI now supports nearly every stage of an attack, from choosing a target to holding a live conversation with them.
More Believable Messages
Polished writing used to signal legitimacy. Now every message has it.
Language models write fluent, error-free text in any language, tone, or level of formality. A message can copy the clipped style of an internal IT notice or the easy familiarity of a longtime vendor.
Models can also match context. An attacker can give a model a real email thread, a press release, or a job posting and get back a message with the right names, project terms, and internal shorthand.
That takes away the cue employees were trained on for years. When a message reads exactly like something a coworker wrote, its wording no longer reveals much.
Automated Reconnaissance for Spear Phishing and Whaling
Spear phishing targets a specific person. Whaling is spear phishing aimed at executives and other high-value people, where one successful message can authorize a large payment or expose sensitive data.
Both used to take hours of manual research per target. Attackers had to read profiles, scan social media, review company bios, and work out reporting lines before writing anything.
AI cuts that work down to minutes. A model can gather a target’s public footprint, including job history, recent posts, speaking engagements, colleagues, and personal interests, and turn it into a usable profile.
Researchers at Harvard and partner institutions tested this directly. Their AI tool built profiles of participants automatically, and those profiles were accurate and useful in 88% of cases.
Each profile then shapes the pretext. A post about a recent conference becomes a follow-up from a supposed fellow attendee. A vacation photo tells the attacker when to pressure an assistant while the executive is hard to reach.
Executives are especially exposed because so much of their professional lives is public. Keynotes, podcast appearances, and press quotes give attackers the personal details they need and the voice samples to go with them.
Social Engineering as a Repeatable Process
AI has also turned social engineering into something close to a production line.
Research, drafting, translation, and follow-up used to depend on a skilled operator. Now they can be chained together and run with little human involvement, so one attacker can run campaigns that once took a team.
In the same Harvard-led study, fully AI-automated spear phishing emails had a 54% click-through rate. That matched emails written by human experts and was far above the 12% rate for generic phishing. The researchers estimated that AI could make phishing up to 50 times more profitable at scale.
That changes the economics. Personalized attacks used to be saved for high-value targets because they were expensive to produce. Once personalization costs almost nothing, every employee is worth a tailored message.
Agentic AI goes a step further. In August 2025, Anthropic reported that a criminal used the Claude Code tool to automate reconnaissance, credential harvesting, and network intrusion against at least 17 organizations, and the AI made both tactical and strategic decisions along the way.
Real-Time Adaptation
Traditional scams followed a script. If the target asked an unexpected question or pushed back, the scammer often stumbled.
AI-driven attacks can adjust as a conversation unfolds. A model can read a reply, pick up on hesitation, and change course within seconds by offering a plausible explanation, shifting tone, or raising the urgency.
Voice adds another layer. Researchers at the University of Illinois Urbana-Champaign built AI voice agents that ran common phone scams against researchers playing the victims. The agents succeeded 36% of the time at an average cost under $0.75 per successful scam.
In one test, the agent asked for the victim’s two-factor authentication code and used it to log in. It then completed about 20 more steps to finish a bank transfer and kept the conversation coherent throughout.
Persuasion Tailored to the Individual
Social engineering has always relied on persuasion. AI makes that persuasion more precise.
Language models learn from huge amounts of human writing, including sales copy, negotiation guides, and customer service scripts. They can use authority, scarcity, reciprocity, and urgency more fluently than most scammers.
They also tailor arguments to the person in front of them. In a study published in Nature Human Behaviour, GPT-4 was given basic demographic details about its debate opponents and was more persuasive than human debaters 64% of the time.
That study used only details such as age, gender, education, and political affiliation. An attacker working from a full social media profile has far more to go on.
Impersonation Across Voice and Video
Voice cloning and deepfake video carry the same realism past text. An attacker can now sound and look like a specific executive, colleague, or family member, which is covered in more detail in the vishing section below.
Why AI-Driven Attacks Are Harder to Stop
Taken together, these attacks are cheaper to launch, harder to spot, and better at handling resistance than the attacks most security training was built around.
Surface-level red flags are no longer dependable. The sections below cover how AI strengthens each attack channel and which behavioral signals still hold up.
What Is Phishing?
Phishing is social engineering delivered through email. It is also the umbrella term for this family of attacks, and smishing and vishing are both forms of phishing.
A phishing email typically pushes the recipient to click a link, open an attachment, or reply with information. Depending on the payload, the outcome can range from stolen credentials to ransomware.
Phishing accounted for 16% of breaches in the 2026 DBIR, level with the prior year. The IC3 received 191,561 phishing and spoofing complaints in 2025.
Modern Phishing Looks Different
AI-written phishing emails rarely contain the grammar mistakes employees were trained to watch for. Many now reference real projects, vendors, or colleagues to appear routine.
Attackers also increasingly aim for account access rather than a single payment. Fake login pages can capture both a password and a multifactor authentication (MFA) code as the victim enters them.
Business email compromise (BEC) remains one of the costliest outcomes. The IC3 attributed more than $3 billion in 2025 losses to BEC.
What Is Smishing?
Smishing is phishing delivered by SMS text or messaging app. It typically impersonates a bank, delivery service, toll authority, or employer and urges the recipient to tap a link or call a number.
Mobile channels tend to work better for attackers than email. Verizon found that in phishing simulations, the median click rate for mobile-centric vectors such as voice and text was 40% higher than email.
Smaller screens, shortened links, and the casual nature of texting all make it harder to examine a message closely before acting.
Smishing is also frequently the opening move in a larger attack. A text may establish a pretext, then direct the target to a phone number where a vishing call completes the scam.
What Is Vishing?
Vishing, short for voice phishing, is social engineering carried out over a phone call or voice message. The attacker poses as a trusted party, such as IT support, a bank, a government agency, or an executive, and pressures the target into sharing information or taking an action.
Vishing has grown sharply. CrowdStrike's 2025 Global Threat Report recorded a 442% increase in vishing between the first and second halves of 2024, fueled by AI-driven phishing and impersonation tactics.
Vishing Attacks Against IT Help Desks
Some of the most damaging recent attacks targeted help desks rather than end users. The attacker calls posing as an employee who is locked out and asks for a password reset or a new MFA device.
In a July 2025 joint advisory update, CISA and the FBI reported that the Scattered Spider group posed as employees to convince help desk staff to reset passwords and transfer MFA to devices the attackers controlled.
The reverse also happens. In a campaign documented by Google Threat Intelligence in June 2025, callers posing as IT support persuaded employees to authorize a malicious connected app in their organization's Salesforce environment, which the attackers then used to steal data.
Voice Cloning and Deepfakes
Voice cloning tools can now produce a convincing replica of a person's voice from a short sample of recorded audio. Leaders who speak publicly on podcasts, webinars, or earnings calls give attackers plenty of material to work with.
In May 2025, the FBI warned of an ongoing campaign using texts and AI-generated voice messages that impersonated senior US officials to gain access to the accounts of their contacts.
Deepfakes now extend to video as well. In early 2024, an employee at engineering firm Arup transferred more than $25 million after joining a video call in which the CFO and other colleagues were digital recreations.
Robocalls and Callback Scams
Not every vishing attack starts with a live caller. Automated calls and voicemails can deliver a scripted message and route anyone who responds to a human operator.
Callback scams flip the approach. A text or email, such as a fake invoice or subscription renewal, prompts the target to call a number where the attacker is waiting.
How Phishing, Smishing, and Vishing Work Together
Today's most effective campaigns rarely rely on a single channel. Attackers combine them to build credibility at each step.
A campaign might open with an email about a security alert, follow with a text containing a verification code, and close with a phone call from someone claiming to be IT support. Each touchpoint makes the next one feel more legitimate.
Defending against one channel alone is not enough. Policies and training need to account for the full chain.
How to Identify a Social Engineering Attempt
Because AI has erased many surface-level tells, the most reliable warning signs are now behavioral rather than cosmetic.
- Unusual urgency. Requests to act immediately, keep something confidential, or bypass normal process are red flags.
- Requests for codes or approvals. Legitimate IT staff and service providers do not need an employee's password or MFA code.
- Changes to payment details. New bank information or a sudden change in how a payment should be sent warrants independent confirmation.
- Channel switching. A conversation that jumps from email to text to phone, or to an unfamiliar platform, may be an attempt to avoid scrutiny.
- Trusting caller ID. Phone numbers can be spoofed, so a familiar name or number on screen is not proof of identity.
How Organizations Can Reduce Their Risk
No single control stops social engineering. The goal is to make each step of an attack harder and to give employees clear permission to slow down.
Verify Through a Separate Channel
Any request involving money, credentials, or access changes should be confirmed using contact information already on file. A callback to a known number defeats most voice and video impersonation.
The FBI also recommends establishing a secret word or phrase with family members to confirm identity. The same principle can apply to high-risk requests within an organization.
Harden Help Desk Procedures
Help desks should require strong identity verification before resetting passwords or enrolling new MFA devices. Manager approval, video verification, or a callback to a number on record all raise the bar for an impersonator.
Adopt Phishing-Resistant MFA
Codes sent by text or approved through push notifications can be stolen or socially engineered. Phishing-resistant methods such as FIDO2 security keys and passkeys are far harder to abuse because they are tied to the legitimate site.
Train for Today's Threats
Security awareness training should reflect what attacks look like now, including AI-written emails, cloned voices, and multichannel campaigns. Simulations that include text and voice scenarios prepare employees for the channels where they are most likely to be caught off guard.
Limit What Attackers Can Learn
Public details about staff, reporting lines, and vendors give attackers the material they need to build a convincing pretext. Reviewing what is shared on websites and social profiles reduces that raw material.
Spotting and Reporting Social Engineering
Most social engineering attempts share a few warning signs, whether they arrive by email, text, phone, or video.
- Pressure to act immediately or keep a request confidential
- Requests for passwords, MFA codes, or access approvals
- Unexpected changes to payment or banking details
- Attempts to move the conversation to a different channel or platform
Reporting quickly limits the damage and helps others avoid the same attack.
- Report internally first. Notify the organization's IT or security team about any suspicious message or call, including cases where a link was already clicked or information was already shared.
- Contact the bank right away. If money was sent, the financial institution may be able to stop or recall the transfer.
- Forward spam texts to 7726. The FTC notes that forwarding texts to 7726 (SPAM) helps wireless providers block similar messages.
- File with federal authorities. Cybercrime can be reported to the FBI's IC3, and fraud and impersonation scams to the FTC at ReportFraud.ftc.gov.
Wrapping Up on Modern Social Engineering
Social engineering has always depended on trust, and AI has made that trust easier to exploit. Phishing, smishing, and vishing now arrive polished, personalized, and often coordinated across channels.
The organizations best positioned to resist them build verification into everyday processes. Clear procedures, modern authentication, and well-informed employees remain the strongest defense against attacks that target people.
Cyberattacks are only going to continue evolving as technology advances. Get ahead of the game by watching Impact’s webinar, Dissecting Cybersecurity Breaches: How They Happen and How to Stop Them.
