Deepfake social engineering is the use of AI-generated or AI-altered voice, video, images, or messages to impersonate someone the target trusts, so that the target sends money, grants access, or reveals information. A joint information sheet from the NSA, FBI, and CISA (September 2023) names three main threats to organizations: executive impersonation that damages the brand, impersonation of leaders and financial staff for financial gain, and impersonation to gain access to networks and sensitive information. The most reliable defense is not learning to spot fakes. It is building verification steps that still work when the fake is perfect.

How the attacks work

The agencies' information sheet explains why this became a mainstream risk. A convincing fake once took a skilled professional days or weeks, but it can now be produced quickly with little or no technical skill, using inexpensive and often free tools. The sheet also points to models that need only a few seconds of reference audio to capture a person's voice. Attackers do not need a perfect fake. They need one that holds up for a few minutes in front of a busy person.

  • Executive impersonation for brand manipulation: a fake video or audio clip of a leader released publicly before it can be disputed, aimed at reputation and, for listed companies, share price
  • Impersonation for financial gain: a fake executive or finance officer authorizes a payment, often by combining voice, video, and text. The sheet describes these schemes as far more common in practice than the brand-focused kind, and as an extension of Business Email Compromise (BEC)
  • Impersonation to gain access: fake candidates in remote job interviews using stolen personal data, or fake customers trying to reach individual accounts
INFOGRAPHIC
What deepfake impersonation is used for
Deepfake
same tools, four different goals
Brand
fake leader clip goes public
Money
fake approval by voice and video
Access
fake candidate or staff
Customers
posing as a customer to get in

What real cases look like

The information sheet describes two attempts from May 2023. In the first, a product line manager was contacted on WhatsApp and invited to a call with someone claiming to be the CEO. The voice sounded right, and the image likely reused an older photo of the CEO. In the second, a caller impersonating an executive used a poor-quality WhatsApp audio call and suggested a Teams meeting where the screen appeared to show the executive in an office. The caller then blamed the bad connection, moved to text, and urged a wire transfer. The target became suspicious and ended the contact.

A larger public case is Arup. According to CNN's reporting, a finance employee in Hong Kong first doubted a request for a secret transaction, then joined a video call with people who appeared to be the chief financial officer and colleagues, all of them deepfake re-creations. The employee sent HK$200 million (about US$25 million) in 15 transactions. Arup confirmed that fake voices and images were used. The investigation was described as ongoing at the time, so treat the details as reported rather than final.

  • Urgency and secrecy: the request must happen now, and nobody else should know
  • Channel switching: a move from email to WhatsApp, or from video to text, away from systems that log and filter
  • A request that moves value: money, credentials, one-time codes, or access
  • Pressure on one person: the target is alone, and the story is built to make a check feel rude
INFOGRAPHIC
A typical deepfake fraud path
01
Collect
public voice and photos
02
Contact
message or call, fake identity
03
Pressure
urgent and secret
04
Switch channel
to chat or text
05
Value moves
payment or codes sent

Why spotting fakes by eye is not a plan

The FBI's December 2024 public service announcement lists visual and audio tells such as distorted hands, unrealistic eyes or teeth, inaccurate shadows, lag, and unnatural movement. They are worth knowing, but they are the weakest line of defense. Quality keeps improving, a bad connection explains artifacts away (as in the May 2023 case), and the information sheet expects a continuing escalation between generation tools and detection tools. It also notes that identity checks during live communication now need liveness testing. Assume any single person can be fooled, and design the process so that no one person's judgment releases money or access.

INFOGRAPHIC
Ways to verify a request: easy or hard to fake
Looks and sounds right
voice cloned from seconds of audio
Sender name, number, chat profile
can be spoofed or hijacked
Questions only the real person knows
helps, but secrets can leak
Call back on a number you hold
channel independent of the attacker
A second approver
no one person can release funds
Easy to fake (riskier)Hard to fake

Controls that work

  • Out-of-band verification: for any request involving money, bank-detail changes, credentials, or access, end the contact and call back on a number you already hold, such as the internal directory, never one supplied in the message. The FBI gives the same advice for unexpected callers
  • Dual approval and limits: require a second approver for payments and payee changes above a threshold, so one convincing call cannot move funds. This is common practice rather than a requirement stated in the agencies' sheet
  • Real-time identity verification: the sheet recommends verification that works during live communication, including liveness testing, plus MFA, one-time codes, or known personal details before sensitive transactions, especially financial ones
  • Phishing-resistant sign-in: a cloned voice cannot talk a passkey out of a login, and one-time codes should never be read out to anyone (the FBI's December 2025 alert warns that scammers ask for them). See Which MFA Is Safer? and Passkeys and FIDO2 Explained
  • Protect executives' media and channels: the sheet suggests extra protection for high-priority people, such as watermarks or content credentials on official media. A short list of the channels your executives really use for approvals is a useful addition
  • Plan, rehearse, and train: run tabletop exercises that include executives, and train staff, including recruiters and finance teams, on how deepfakes are used and where to report them

If you suspect a deepfake

Stop the transaction or the conversation without confronting the caller, then verify through a channel you already trust. Treat it as a security incident, not an embarrassment, because the same actor often tries several employees. The information sheet's basic steps for examining suspicious media are below.

  • Preserve before analyzing: copy the media, then hash the original and the copy to confirm they match
  • Check the source before drawing conclusions: who sent it, and through which account or channel
  • Look and listen first, then use tools: try reverse image search and metadata examination. Intact metadata supports authenticity, but stripped metadata may only mean a platform removed it, so it proves nothing alone
  • Report: the sheet asks organizations to report malicious deepfakes to the FBI, and the FBI's alerts point to IC3 for incidents. If money has moved, tell your bank immediately
INFOGRAPHIC
What to do when you suspect a deepfake
1
Pause
hold the payment or conversation
2
Verify out-of-band
call back on a number you hold
3
Preserve evidence
copy the file, hash original and copy
4
Report
security team, bank, IC3
5
Fix the process
close the gap the attack used

Provenance works in the opposite direction from detection. Instead of guessing whether media is fake, standards such as C2PA, which the sheet describes as an effort launched in 2021 to certify the provenance of media, let creators attach verifiable history to authentic content. This helps prove that your official statements are real, but it does not stop an attacker from making a fake without credentials, so it complements verification procedures rather than replacing them.

Common mistakes

  • Relying on detection tools or staff instincts alone: both fail sometimes, and a false sense of security is worse than a known gap
  • Verifying inside the same channel: replying "is this you?" on the thread or call the attacker controls proves nothing
  • Exempting senior people from the process: the most convincing fake is the CEO's, and staff will not challenge it unless the policy says everyone follows the same steps
  • Training once: attack formats change, so refresh the examples yearly and after incidents
  • Ignoring non-finance targets: help desk, HR, recruiters, and customer service handle access, identity, and accounts too
  • Hiding near-misses: staff who report a failed attempt protect others. Make reporting easy and blame-free

Start today

  • Write a one-page rule: any request for money, bank-detail changes, credentials, or access over voice, video, or chat must be confirmed by call-back on a known number
  • Add a second approver for payments and payee changes above a threshold you choose
  • List your high-risk workflows: payment release, password reset and MFA enrollment, vendor changes, hiring, and customer account changes
  • Run a 30-minute tabletop: a fake CEO call to a finance employee late on a Friday. See what happens
  • Check your sign-in methods: move high-risk accounts from codes that can be read out to phishing-resistant methods

Not sure where your gaps are? Try the self-assessment for a prioritized view. For the identity side of these attacks, see also ITDR: Identity Threat Detection and Response.

Forging the sender's name is one way a deception starts. An organization can make forging its own domain harder with the setup described in DMARC, SPF and DKIM: How to Set Them Up Correctly.

Attackers already send fake payment requests by email, and faked voices and faces make a phone confirmation insufficient on its own. For the email-side layers, see Email Security: Layered Defense Against Phishing and BEC.

Training staff to resist deception, by email and by fake voice alike, works best when it is blame-free and measures reporting. See Security Awareness and Phishing Simulation That Works.

References and further reading