AI Deepfake Incidents: Documented Cases, Patterns and What Regulators Say

What counts as an AI deepfake incident?

An AI deepfake incident is any event in which synthetic or AI-altered media is used to deceive a person or a system and causes measurable harm: money lost, an account opened under a false identity, a market moved, or a person abused. The media can be a cloned voice on a phone call, a face swapped into a live video meeting, an AI-generated selfie fed into an onboarding flow, or a fabricated image shared on social media.

The FBI's public service announcement of 3 December 2024 (I-120324-PSA) gives a useful catalogue. It lists AI-generated images used as profile photos for romance and investment scams, forged driver's licences and bank credentials, cloned voices of relatives demanding ransom, cloned account-holder voices used to get into bank accounts, and real-time video used to pose as executives or police. Most incidents combine one of these media types with an old fraud script.

Where deepfakes stand right now

Deepfakes are now an ordinary tool of fraud and abuse rather than a novelty. Three official data points show the scale. NCMEC counted 4,700 CyberTipline reports linked to generative AI in 2023, 67,000 in 2024 and more than 400,000 in 2025. The FBI's Internet Crime Complaint Center recorded a record $16.6 billion in reported losses across 859,532 complaints in 2024, a 33% rise on 2023, with business email compromise alone at $2.77 billion. And MAS wrote in September 2025 that deepfake incidents involving impersonation, falsified documents and fraudulent transactions had already caused losses to organisations worldwide.

IC3 does not break out how many of its complaints involved synthetic media, so the true deepfake share of fraud losses is unknown. What the record does show is a shift in where deepfakes appear: from viral videos of celebrities toward private, targeted channels such as video calls, onboarding selfies and voice notes, where there is no audience to spot the fake and no time to check.

What did MAS Singapore say about deepfake risks?

The Monetary Authority of Singapore published the information paper Cyber Risks Associated with Deepfakes on 18 September 2025 (MAS/TCRS/2025/06). It says deepfakes threaten financial institutions' biometric authentication and their email, video and voice communication, and it maps the harm onto market, cyber, fraud, regulatory and reputational risk. It builds on MAS's July 2024 paper on generative AI cyber risks.

The paper focuses on three areas. For biometric bypass, it recommends tamper and authenticity checks on documents and content at onboarding, liveness detection, regular testing against deepfake samples from several generation methods, and endpoint protection against injection attacks. For impersonation scams, it recommends staff and customer awareness campaigns, including deepfake simulation exercises, and extra verification for high-risk transactions. For misinformation, it recommends monitoring for brand and executive impersonation and adding deepfake scenarios to incident response playbooks. MAS also notes that real-time deepfake detection built into communication channels can warn users during calls.

Documented deepfake incidents, 2023 to 2025

The cases below are drawn from the incident sections of the MAS paper, which cites police statements and press reports for each. They show the same three patterns repeating across countries and years.

Two details stand out. In the Hong Kong video-conference case, the fraudsters built the fake CFO and colleagues from publicly available video footage. In the Indonesian lending case, a cybersecurity firm's analysis found the attackers used virtual camera software to feed pre-prepared deepfake photos into the KYC check, so the system treated recorded media as a live capture.

What the incident record teaches defenders

Attackers rarely need technical brilliance. They need source material, a channel and a pretext. Public conference videos, earnings calls and social media posts supply executives' faces and voices. Messaging apps provide the first contact, and a move to a video call adds credibility. The Singapore case followed this path: the victim was first approached on WhatsApp, then invited to a Zoom meeting with a fake CEO.

The same case shows what stops a loss. The finance director became suspicious when a second transfer of US$1.4 million was requested, alerted the corporate bank, and the Singapore Police Force's Anti-Scam Centre worked with Hong Kong authorities to recover the first payment. A pause, a question and a fast call to the bank did more than any visual inspection of the video.

Onboarding incidents teach a different lesson. When fakes enter through a virtual camera, a liveness check that only looks at the picture can be fooled by pre-recorded or generated media. MAS's call for endpoint protection against injection attacks and testing with deepfake samples from multiple generators reflects that gap.

Where detection helps after an incident, and where it falls short

Detection helps most when there is a file to examine: a recorded call, a saved meeting, an onboarding selfie or a viral image. It can show which frames or regions look manipulated, support an insurance claim or police report, and tell a platform whether a takedown request concerns synthetic media. It also helps a firm find sibling cases, for example other onboarding sessions that used the same generated face.

It helps less when nothing was recorded, when media has been compressed and re-shared many times, or when the deception was mostly in the script rather than the pixels. A detector reports a probability, and analysts still need context: who sent the media, through which channel, and what was asked for. Treat a detection result as one strand of evidence alongside account logs, payment records and witness statements.

Investigating deepfake incidents with Neural Defend

NeuroDetect is Neural Defend's forensic workspace for this after-the-fact work. Analysts upload authorized images, audio, video or documents into a case; each analysis keeps its model and version context; and detailed views, including explainability heatmaps and attention views, drill down to regions, frames and audio segments. An LLM-assisted summary explains findings in plain language with links back to the evidence, the analyst makes the final call, and the downloadable forensic report records observations, uncertainty, provenance, model context and analyst notes.

For prevention at the points MAS highlights, onboarding and high-risk transactions, NeuroVerify checks a face image, video or audio in one API call across six signals and returns a single risk score with signal-level evidence. Both products run on ARCAS 1B, Neural Defend's own detection model; the ARCAS paper reports pooled accuracy of 99.99% on GenImage and 97.65% on OpenFake. Deployment can be managed cloud, private cloud, on-premises or air-gapped, and Neural Defend holds ISO/IEC 27001, ISO/IEC 42001 and SOC 2 Type II certifications.