Research and resources

The evidence behind AI voice scams, cited.

RefugeID is built on published evidence rather than headlines. Everything on this page comes from the AI Safety Research Group's risk register and the primary documents behind it: regulators, the FBI, peer-reviewed work and named newsrooms. Nothing here is our own research, and every link opens the original.

09
Risk categories tracked
44
Sources cited
00
Synthetic cases

Risk register

Nine categories of AI-enabled harm, ranked.

Ranked by severity, exposure and how well the capability is documented. Each entry states what its evidence does not establish, which is the part worth reading.

Rank 01Critical

Synthesized-voice family extortion

Fake and take

A cloned parent or child voice stages an emergency and demands immediate payment. The family member is always safe.

Under 30 seconds of audio clones a voice

Read the entry
Rank 02Critical

Deepfake-authorised payment fraud

Executive impersonation

Financial controls defeated by a face and a voice synthesized in real time.

$25.6M largest documented single loss

Read the entry
Rank 03Severe

AI-automated spear phishing

Reconnaissance at scale

Reconnaissance and personalisation automated to a cost that makes targeted attacks routine.

54% click-through in a four-arm trial

Read the entry
Rank 04Severe

Synthetic identity and KYC bypass

Onboarding defeated

Liveness checks defeated by injected, synthesized video at the point of account opening.

$27.3B US identity fraud losses, 2025

Read the entry
Rank 05High

Agentic vulnerability exploitation

Autonomous offensive capability

Agents that find and exploit software flaws autonomously, at negligible cost per attempt.

73.3% of prepared targets exploited

Read the entry
Rank 06High

Non-consensual synthetic imagery

Image-based sexual abuse

Sexualised forgeries generated at scale, overwhelmingly depicting women and girls.

98% of deepfake video is pornography

Read the entry
Rank 07High

Official impersonation and political audio

Spoofed authority

Authority spoofed at the scale of a phone bank, with enforcement arriving afterwards.

Illegal in the US since February 2024

Read the entry
Rank 08Watch

Spoofing of voice authentication

Speaker verification

Voiceprints treated as credentials, against attack families that are still unsolved.

32 attack algorithms in the current benchmark

Read the entry
Rank 09Watch

Industrialised consumer fraud

Generative media at volume

The largest losses are not novel techniques. They are old ones, made cheap to run at volume.

$12.5B reported US fraud losses, 2024

Read the entry

Primary sources

Read the documents yourself.

Regulators, law enforcement, technical research and reporting on documented cases. Every link goes to the original publisher, not a summary of it.

How to read this

A number you can check beats a number that sounds alarming.

The index is not an incidence count

It orders risk by harm, exposure and documented capability. It never claims the top category harms the most people.

Vendor percentages are direction of travel

Figures measured across one company’s customer base are not a rate of victimisation, and we do not present them as one.

Unknowns stay unknown

No public dataset isolates voice-cloning family extortion, so the aggregate is reported and the subtype labelled undercounted.

Methodology, sourcing standards and limits

RefugeID

Put the research to work.

RefugeID screens every call for AI-generated voices and cloned family members before anyone picks up. $8 per month, per line, and 25% off your first year when you pre-order and pay the year up front.

By NausicaAI. No spam, just protection.

RefugeID

Know who is calling before anyone picks up. RefugeID is AI voice detection and call screening that flags cloned voices and phone scams for the families who need it most.

Company

  • NausicaAI
  • Founded for families
  • United States

© 2026 NausicaAI. RefugeID is a NausicaAI company.

No recordings kept. No voices sold.