Back to Blog

Deepfakes in 2026: how they evolved, how to spot them, and where they are headed

Deepfakes in 2026: how they evolved, how to spot them, and where they are headed

Overview

Deepfake fraud is now one of the costliest cybersecurity categories. The FBI logged more than 22,000 AI-related complaints in 2025 with close to $900 million in confirmed losses. Multiple industry reports also show rapid adoption of deepfake tactics across businesses, including:

What changed is not just the volume of attacks. The cost and complexity for attackers fell sharply. A usable voice clone can be created from only a few seconds of audio. Face swaps run on consumer hardware. Real-time, multi-person video deepfakes have moved from demos to active criminal operations.

This report explains what deepfakes are, what changed in 2026, how to spot them during a live interaction, how to defend at the protocol and technology level, and what to expect next.

What is a deepfake, and why does 2026 matter?

A deepfake is synthetic media generated by AI that depicts a person saying or doing something they did not say or do. In practice, “deepfake” covers several attack types:

  • Voice clones used in phone calls and voicemail
  • Face swaps used in video calls
  • Fully generated video used in scams and disinformation
  • Manipulated images used for document forgery and identity bypass

The underlying research has existed for years, but three forces converged in 2025 and 2026:

  1. Open and commercial generative models reached consistently high quality.
  2. Consumer hardware became fast enough for real-time generation.
  3. Tools emerged that packaged the capability into point-and-click workflows.

The World Economic Forum’s January 2026 Cybercrime Atlas documented face-swapping and camera-injection tools capable of bypassing standard biometric onboarding checks.

The key implication is economic. Attackers no longer need advanced technical skill. Deepfake capability is increasingly bought as a service, bundled with scripts, target lists, and pre-trained voice and face models.

By 2026, enterprise surveys consistently rank deepfake fraud as a top-tier concern, and insurers are beginning to price deepfake exposure explicitly in cyber policies.

What evolved in deepfake attacks in 2026?

Five shifts defined the year.

1) From pre-recorded to real-time synthetic video in live calls

High-profile incidents demonstrate the pattern: the Arup case in Hong Kong ($25.6 million, January 2024) and the Singapore Zoom incident ($499,000, March 2025) involved real-time, multi-person deepfake meetings.

By 2026, the tooling required to run multi-participant synthetic calls is widely available. Confirmed cases have been reported across multiple regions.

2) Rapid growth in voice-clone fraud targeting contact centers

Researchers reported more than a 1,300 percent increase in contact-center deepfake fraud attempts during 2025, with voice-based phishing becoming a large share of incident response work in several markets.

Two drivers explain the shift:

  • Voice-clone generation costs are near zero.
  • Many contact-center verification flows were not built to detect synthetic audio.

3) Camera injection against biometric identity verification

Attackers increasingly bypass “presentation attack” defenses by feeding synthetic video directly into the device or app verification pipeline using virtual cameras, manipulated drivers, or modified mobile apps.

Threat reporting showed a 741% annual increase in biometric injection attacks, including a surge on iOS. Some institutions have publicly acknowledged account-opening fraud at scale, including a case where attackers opened 46 fraudulent accounts before detection.

4) Multi-channel campaigns: voice + video + documents

A common 2026 pattern combines:

  • Synthetic voice or a synthetic meeting to establish a pretext
  • Follow-on messages to create urgency and secrecy
  • Forged supporting documents such as invoices, wire instructions, or legal opinions

The Ferrari impersonation attempt (July 2024) is an early example of this coordinated approach.

5) Early agent-to-agent impersonation inside automation workflows

As organizations deploy AI agents for procurement, support, and finance operations, new identity threats emerge: spoofing, replay, or impersonation of an internal agent in an authorization chain. This category is still small by volume in 2026, but it is growing as agentic deployments expand.

How to spot a deepfake in 2026

Point-of-contact detection still matters because many deepfakes continue to produce artifacts. Practical detection breaks into three layers.

Visual indicators (video)

Look for:

  • Lighting and shadows that do not match the environment
  • Overly smooth skin texture
  • Flicker or warping around the jawline or hairline during movement
  • Unnatural blinking or eye tracking
  • Teeth that lose detail when the mouth opens
  • Lip-sync drift, especially on “p” and “b” sounds

Audio indicators (voice)

Listen for:

  • A “flat” or compressed voice that lacks micro-variation
  • Background noise that cuts in and out rather than staying consistent
  • Reverb or room acoustics that do not match what the camera shows
  • Speech that is technically clean but emotionally inert

The most reliable method: behavioral verification

Visual and audio cues are useful, but behavioral checks work best in the moment.

Use prompts the attacker cannot pre-script, for example:

  • A specific shared memory or a detail from a recent in-person conversation
  • A request to perform a random physical action (turn sideways, cover one eye)
  • A request to read a phrase you provide on the spot

Real-time deepfakes still struggle with profile views, improvisation, and out-of-band physical requests. Hesitation, freezing, or sudden topic changes are strong signals to escalate verification.

How to stay safe from deepfake attacks

Effective defenses are layered: personal habits, organizational protocols, and technical controls. No single layer is sufficient.

Personal protections

Use a pre-agreed safe word with family and high-trust contacts for any suspicious call involving money, identity, or emergencies.[1]

Store the safe word in a password manager. Use it the first time an urgent request feels off.

Organizational protocols (highest ROI)

Implement an out-of-band callback rule for financial or credential-related requests:

  • Any payment, wire, vendor change, or credential reset initiated by video, voice, or chat must be confirmed by calling back on a number from the internal directory.
  • Do not use the number provided in the original message.

Add friction for high-risk workflows:

  • Two-person approval above a defined threshold
  • A cooling-off period before wires to newly added accounts
  • Known-channel or in-person verification for “urgent” and “confidential” requests
  • Quarterly tabletop exercises simulating a deepfake executive call against the real approval workflow

Technology layer

Adopt controls that reduce reliance on human perception:

  • Provenance verification where available (for example, C2PA)
  • Active liveness for biometric onboarding (randomized user actions, challenge-response)
  • Behavioral analytics for call and session patterns
  • Multimodal deepfake detection integrated into conferencing and contact-center workflows

The practical rule is simple: the cost of slowing down a legitimate request is usually lower than the cost of approving a synthetic one.

Outlook: 2027 and beyond

Based on field work across identity, payments, and agentic AI, these are the most likely developments through 2028.

  1. Visual artifacts will largely disappear. Within 12 to 18 months, blink anomalies, edge flicker, and lip-sync drift will become less common. Behavioral verification will matter more than “spot the artifact.”
  2. Provenance moves toward regulation. Standards like C2PA are likely to shift from voluntary adoption to requirements in some jurisdictions. The EU has signaled expansion of the AI Act to synthetic media labeling, with parallel efforts elsewhere.
  3. “Voice as a password” will fade. Voice authentication alone will be too easy to spoof. Contact centers and banks will increasingly require liveness, behavior, or knowledge-based checks.
  4. Insurance will enforce controls. Underwriters are beginning to price deepfake exposure explicitly. Expect policy terms to require protocols such as callbacks, dual approvals, and exercises.
  5. Agent identity becomes a security domain. Organizations will need cryptographic identity for agents, signed action chains, and attestation to prevent agent-to-agent spoofing.
  6. Election cycles will accelerate disinformation tooling. Industrial-scale deepfake operations will push platforms and regulators to set new governance precedents.

The bottom line: trust in remote audio and video for high-value decisions will continue to erode. Organizations will need either in-person verification for critical actions or cryptographic equivalents that do not depend on humans recognizing a face or voice.