Back to articles

Deepfake and synthetic media detection: what security teams need to know

The short answer

For a security team, deepfake detection is two jobs. The first is checking media that reaches you, such as a video call or voice message, with verification steps that do not rely on how real it looks. The second is finding synthetic impersonations of your executives and brand that circulate outside your walls, and removing them. Detector scores help, but process controls and external monitoring are what stop the fraud.

Key takeaways
  • ↗An Arup employee in Hong Kong sent HK$200 million (about US$25 million) after a video call with a deepfaked CFO and colleagues.
  • ↗The FBI has warned since May 2025 about AI-generated voice messages impersonating senior US officials, often followed by a move to encrypted messaging apps.
  • ↗Content Credentials show where media came from and whether it was edited. They do not prove that unlabeled media is fake.
  • ↗Verification protocols for payments and access requests stop more fraud than trying to spot a fake by eye.

What changed for security teams

Deepfakes moved from novelty to fraud tool. In early 2024, an employee at engineering firm Arup joined a video call with what appeared to be the company's CFO and other colleagues. Every other participant was synthetic. The employee made 15 transfers to five bank accounts, totaling HK$200 million, or about US$25 million.

Voice is now as easy to clone as a face. In May 2025 the FBI warned that criminals were sending AI-generated voice messages that impersonated senior US officials. Its December 2025 update said the activity dates back to 2023, that targets were moved to apps such as Signal, Telegram, and WhatsApp, and that attackers asked for authentication codes.

Synthetic media also targets onboarding. In November 2024, FinCEN alerted financial institutions to fraud schemes using deepfake media, including AI-altered identity documents built to pass verification checks.

Four kinds of synthetic media attacks

Live calls

Real-time video and voice impersonation

An attacker joins a meeting or calls as an executive to approve a payment, a vendor change, or an access request.

Messages

Cloned voice notes and voicemail

A short voice message builds trust, then moves the target to an encrypted app where the real request is made.

Public content

Fake endorsements and scam ads

Synthetic videos of executives promote fake investments, giveaways, or apps on social platforms and ad networks.

Onboarding

Synthetic identity documents

AI-generated or altered IDs and selfies try to pass know-your-customer and hiring checks.

Why detection tools alone fall short

Media forensics tools score how likely a file is to be synthetic. Those scores are probabilities, they vary with compression and quality, and generators keep improving. A real-time call gives an employee seconds, not a forensic workflow.

Provenance standards help from the other direction. The C2PA's Content Credentials attach a signed record of where media came from and how it was edited, so tampering becomes visible. The C2PA is explicit that credentials do not judge whether content is true, and most media online carries none. Missing credentials are not proof of a fake.

Design controls so that a perfect fake still fails. If a payment needs a callback to a known number, a convincing video call is not enough to move money.

A layered defense

LayerWhat it doesUsual owner
Payment and approval controlsOut-of-band callbacks on known numbers, dual approval, and holds on new payeesFinance and treasury
Identity verificationLiveness checks, document checks, and review of FinCEN red flagsFraud, KYC, and HR
ProvenanceContent Credentials on official media so your real content is verifiableCommunications and brand
External monitoring and takedownFinds synthetic impersonations of executives and the brand and removes themSecurity, brand protection, executive protection
TrainingTeaches staff that urgency plus a new channel is the warning sign, not visual glitchesSecurity awareness

Build a verification protocol for high-risk requests

01Verify payment, vendor, and access requests through a second channel you already trust
02Call back on a number from your directory, never one given in the request
03Require two approvers for transfers above a set amount
04Agree code words or challenge questions with executives and their assistants
05Treat any move to a personal or encrypted app as a reason to pause
06Never share authentication codes, even with a caller who sounds like leadership

Monitoring for deepfakes of your executives and brand

01

Map who is exposed

List executives, spokespeople, and brands most likely to be cloned, with names, aliases, and reference photos and voice samples.

02

Watch where fakes spread

Monitor social platforms, video sites, ad libraries, app stores, and messaging channels, not only search results.

03

Read the media, not just the text

Many fakes never name the person in the caption. Detection has to analyze faces, voices, and video frames.

04

Connect it to the campaign

Link each fake to the accounts, ads, domains, and payment details behind it, so removal hits the whole operation.

05

Remove and document

File takedowns with each platform, confirm removal, and keep evidence for legal, communications, and law enforcement.

How Outtake helps

Outtake executive protection

Find synthetic impersonations and take them down

Outtake's Entity Registry holds your executives and brands with aliases and reference images. Agentic Search reads text, images, video, and audio across social platforms, ad networks, marketplaces, and messaging channels, so it can find a deepfake that never names its target.

  • ↗Detection covers deepfakes, cloned voices, impersonated profiles, and doxxing
  • ↗Executive protection checks for new threats about every 15 minutes
  • ↗Threat Graph links each fake to the campaign behind it
  • ↗Remediation Agents file takedowns and re-check until items are confirmed down

Read our executive impersonation guide or learn more about the Outtake platform.

Frequently asked questions

What is synthetic media?

Synthetic media is image, video, audio, or text created or altered with AI. Deepfakes are synthetic media that impersonate a real person's face or voice.

Can deepfakes be detected reliably?

Not with certainty. Detection tools give probability scores that vary with file quality and new generation methods. Pair detection with verification steps, provenance checks, and monitoring for fakes that circulate publicly.

How do I protect executives from deepfake impersonation?

Verify high-risk requests through a known second channel, agree challenge phrases, limit public voice and video samples where practical, and monitor social platforms, ads, and messaging apps for synthetic impersonations so they can be removed quickly.

Do Content Credentials stop deepfakes?

No. Content Credentials record where media came from and how it was edited, which helps verify your own content. They do not judge whether content is true, and media without credentials is not proof of a fake.

Who should own deepfake risk?

It is shared. Finance owns payment controls, fraud and HR own identity checks, communications owns official media, and security or executive protection owns monitoring and takedown of public impersonations.

Related resources

Sources