In early 2024, a finance employee at the engineering firm Arup joined a routine video call. The chief financial officer was there, so were several colleagues. Everyone looked and sounded exactly right. Over the following days the employee made 15 transfers totalling about 25 million US dollars. Every person on that call except the employee was a deepfake3.
That case is not an outlier any more, it is a preview. Deloitte projects that generative-AI-enabled fraud losses in the United States could reach about 40 billion dollars by 2027, up from 12.3 billion in 20232. The FBI logged 21,442 business email compromise complaints and 2.77 billion dollars in losses in 2024 alone1. The tools that create a convincing fake voice or face are now cheap, fast and in everyone’s hands.
So every vendor now sells “deepfake detection”, and the market is a wall of near-identical claims. This guide cuts through it. We name the real tools, group them by the job they actually do, give honest trade-offs and pricing where it is public, and show where each one fits. Superkind appears too, but as one option for one specific job, not as the answer to every row. It is written for the operations, finance, risk or IT leader who has to choose, and who is tired of comparison pages where the company publishing them wins every category.
TL;DR
Deepfake defense is four jobs, not one - real-time media detection, voice and liveness verification, BEC and email impersonation defense, and out-of-band process verification. No single tool wins all four.
Media detection leaders - Reality Defender, Sensity AI, Intel FakeCatcher and Hive, each with a different modality and sweet spot.
Voice, liveness and email leaders - Pindrop and iProov for voice and liveness, plus Abnormal Security, Proofpoint, Microsoft Defender for Office 365 and Mimecast for business email compromise.
The gap every detector leaves open - a human still receives the request, feels the pressure and moves the money. Detection narrows the odds, process stops the loss.
Where Superkind fits - the verified-request backstop that knows your real approval chains and payment processes, so a request that does not match your process is caught before money moves.
The 2026 Deepfake Threat: Why CEO Fraud Got Cheap
The reason CEO fraud exploded is not that people got more gullible. It is that the cost of a convincing impersonation collapsed. A cloned voice used to need a studio and a specialist. Now it needs a few seconds of audio from a conference talk or a voicemail and a consumer tool.
- Deepfakes went from novelty to weapon - in the 2024 Arup case, a finance worker paid out roughly 25 million dollars after a video call in which the CFO and every other participant was AI-generated3.
- Voice phishing is surging - CrowdStrike recorded a 442 percent rise in vishing between the first and second half of 2024, driven by social-engineering operations that increasingly use AI-generated audio4.
- Voice-clone attempts spiked - Pindrop reported a sharp year-over-year rise in deepfake voice fraud attempts against contact centres, the channel where a cloned executive voice does the most damage5.
- Injection attacks are the new front - iProov reported injection attacks, where a forged feed is pushed straight into the camera channel to defeat identity checks, rising around 783 percent in 20246.
- BEC is the quiet giant - the FBI logged 21,442 business email compromise complaints and 2.77 billion dollars in reported US losses in 2024, with nearly 8.5 billion lost to BEC across 2022 to 20241.
- The curve is steep - Deloitte models US fraud losses climbing to about 40 billion dollars by 2027 as generative AI scales, a compound annual growth rate near 32 percent2.
Key Data Point
A 2025 Gartner survey of cybersecurity leaders found that a majority had faced at least one deepfake attack in the prior year, with audio-call incidents the most common and video-call incidents close behind21. Deepfake fraud is no longer a tail risk reserved for banks. It is a mainstream attack on any company that moves money on instruction.
This is the backdrop for the tool market. Every vendor is responding to the same pressure, which is why the marketing converges. The differences that matter are underneath the claims, in what each category of tool actually does.
“These artificially generated images of real people’s faces, known as deepfakes, can be used by malicious actors to undermine biometric authentication or render it inefficient.”
- Akif Khan, VP Analyst at Gartner16
The 4 Jobs of Deepfake Defense
The single biggest buying mistake is treating “deepfake detection” as one purchase. It is four distinct jobs, each with its own leaders, its own pricing model and its own failure modes. Map your actual exposure to the job first, then shortlist tools.
- Real-time media detection - scoring audio, video, images and documents to decide whether the media itself is synthetic, in onboarding, high-risk video and content review.
- Voice and liveness verification - proving the person on a call or at a camera is a live human and not a clone or an injected feed, in the call centre and at identity moments.
- BEC and email impersonation defense - catching the spoofed, compromised or look-alike email that carries most CEO-fraud instructions before a human ever acts.
- Out-of-band process verification - confirming the request itself through a second channel and against your real approval rules, so even a perfect fake fails at the payment step.
| Job | Core question | Typical leaders | Pricing model |
|---|---|---|---|
| Real-time media detection | Is this audio or video synthetic? | Reality Defender, Sensity, Intel FakeCatcher, Hive | Enterprise licence / per call |
| Voice & liveness verification | Is this a live human or a clone? | Pindrop, iProov | Per call / per check / per seat |
| BEC & email defense | Is this email impersonation? | Abnormal, Proofpoint, Microsoft, Mimecast | Per mailbox |
| Process verification | Does this request match our real process? | Callback rules, approval controls, Superkind | Process / platform |
Most companies need three or four of these layers, not one. The hard part is not picking a winner in each row, it is making them work together, which the rest of this guide covers category by category.
Real-Time Deepfake Detection Tools
This is the job most people mean when they say “deepfake detection”: a model that looks at audio, video, an image or a document and scores how likely it is to be synthetic. The leaders differ mainly by modality and by whether they work on a live stream or on uploaded files.
The tools, honestly
- Reality Defender - multimodal detection across audio, video, images and documents, with real-time risk scoring through an API and web app, and plugins for video-conferencing tools. It has deployed real-time detection for large banks, media and governments, and is the strongest general-purpose choice for enterprise deepfake defense9.
- Sensity AI - visual-first detection built for video and image forensics, with broad coverage of generation techniques. A strong fit when the job is investigating or reviewing media rather than scoring a live call10.
- Intel FakeCatcher - a real-time video detector that reads photoplethysmography, the tiny colour changes in a face caused by blood flow, from 32 facial regions, and returns a result in milliseconds. Intel claims about 96 percent accuracy on its test data and up to 72 concurrent streams. Novel and server-based rather than a plug-and-play SaaS8.
- Hive - image and content-focused detection, strong for platforms moderating AI-generated media at scale rather than for executive-payment flows. Useful as a content layer, not a call-centre tool.
- The honest caveat - every accuracy figure in this category is a vendor claim on a vendor benchmark, usually without a stated false-positive rate, and accuracy drops against attacks the model has never seen. MITRE ATLAS has documented live deepfake injection evading both passive and active liveness checks11. Run a proof of concept on your own media.
| Tool | Modality | Best for | Honest note |
|---|---|---|---|
| Reality Defender | Audio, video, image, document | Enterprise, real-time, banks | Enterprise contract and integration |
| Sensity AI | Video, image (visual-first) | Forensics and media review | Audio is a secondary layer |
| Intel FakeCatcher | Video (blood-flow signal) | Real-time video, on-prem | Server-based, vendor accuracy claim |
| Hive | Image and content | Platform content moderation | Not built for payment flows |
“What makes us human? The answer is in our blood.”
- Ilke Demir, Senior Staff Research Scientist at Intel Labs8
Media detection is a necessary layer, but on its own it is a single sensor on one channel. The moment the attack moves from a recorded file to a live voice on a phone line, you need the next job.
Voice and Liveness Verification Tools
This is the job that defends the call centre and the identity moment: proving the voice on the line or the face at the camera is a live human and not a clone or an injected feed. It is where most CEO-fraud and help-desk attacks actually land.
The tools, honestly
- Pindrop - the leader in audio deepfake detection for contact centres. Its Pulse liveness scoring examines signal-level features like frequency shifts and spectral distortions, updates the score during a call rather than only at the start, and claims around 99 percent accuracy against known voice-cloning models and 90 percent against new ones, both at under 1 percent false positives on its own benchmarks. It won the large-organisation category of the FTC voice cloning challenge5,20.
- iProov - the leader in face liveness and injection-attack detection, using a controlled illumination sequence to confirm a genuine live person at the moment of capture. In 2026 its Dynamic Liveness became the first product to meet the CEN/TS 18099 “High” specification for injection attack detection in an extended independent evaluation. Widely used in European eID programmes7.
- Voice biometrics as a layer - established voice-biometric engines authenticate a caller against an enrolled voiceprint. Useful for returning customers, but a cloned voice can target exactly this control, which is why liveness and anti-spoofing matter more than matching alone.
- The honest caveat - liveness is a moving target. A MITRE ATLAS case study found live deepfake injection evading both passive and active liveness in a red-team exercise, and Gartner predicted that by 2026 deepfakes would push 30 percent of enterprises to stop treating face biometrics as reliable in isolation11,16.
| Tool | Signal | Best for | Honest note |
|---|---|---|---|
| Pindrop | Audio anomalies, liveness | Call-centre voice fraud | Accuracy figures are vendor benchmarks |
| iProov | Face liveness, injection detection | Identity and eID moments | Strong standards fit, capture-moment only |
| Voice biometrics | Voiceprint match | Returning-caller authentication | Clones target this control directly |
Specialist Deepfake Detector vs Built-In Liveness
Specialist detector (Pindrop, Reality Defender, iProov)
- ✓ Purpose-built models - tuned specifically for synthetic media and injection attacks
- ✓ Independent evaluation - some meet published standards like CEN/TS 18099
- ✓ Real-time scoring - works during a live call, not just at onboarding
- ✗ Another contract and integration - a separate layer to buy and run
Built-in liveness in an IDV product
- ✓ One vendor - bundled with identity verification you may already run
- ✓ Covers onboarding - fine for the account-opening moment
- ✗ Weak on live calls - rarely protects a payment-approval conversation
- ✗ Varies widely - depth of anti-injection defense is inconsistent between vendors
Worried a perfect fake could move your money?
Book a 30-minute call. We will map where a verified-request backstop fits your approval chains.

BEC and Email Impersonation Defense Tools
Before a deepfake voice ever reaches an employee, most CEO-fraud starts with an email: a spoofed address, a compromised supplier account, or a look-alike domain carrying the urgent instruction. Catching that email is the cheapest, highest-leverage layer most companies already half-own.
The tools, honestly
- Abnormal Security - the behavioral-AI specialist for business email compromise. It baselines normal behaviour for each employee and vendor relationship and flags deviations, deploys through the Microsoft or Google API without changing your mail routing, and is strongest exactly where gateways are weakest, on text-only impersonation with no malicious link12.
- Proofpoint - the enterprise default, with the largest installed base, deep threat intelligence and strong remediation. Heavier to run as a gateway architecture, and buyers most often leave over renewal pricing and operational weight rather than detection quality13.
- Microsoft Defender for Office 365 - the natural fit if you are already on Microsoft 365, at roughly 2 US dollars per user per month for Plan 1 and 5 for Plan 2. Native and cost-effective, though independent tests often rate it a step behind the specialists on the most targeted attacks, and configuration matters a lot14.
- Mimecast - chosen mainly where archiving, continuity and sandboxing are the gating requirement, with solid impersonation controls layered on top. A strong fit for compliance-heavy sectors15.
- The honest caveat - the strongest architecture most guides recommend is not one product but two: a gateway or native platform for commodity filtering, plus an API-based behavioral tool for the targeted BEC that slips past rules13.
| Tool | Approach | Best for | Honest note |
|---|---|---|---|
| Abnormal Security | Behavioral AI, API-based | Targeted BEC and impersonation | Quote-based, best as a layer |
| Proofpoint | Gateway, threat intel | Large enterprise, broad coverage | Heavier ops, renewal pricing |
| Microsoft Defender for O365 | Native to Microsoft 365 | Microsoft shops, cost focus | Trails specialists on targeted attacks |
| Mimecast | Gateway + archive + sandbox | Compliance and continuity | Picked for archiving as much as detection |
Where email defense stops short
Even the best email filter cannot stop a request that arrives by phone, WhatsApp or a live video call, which is exactly where deepfake CEO-fraud is heading. Email defense shrinks the attack surface, but it does not verify the request itself. That is the gap the last two jobs close.
The Verified-Request Process: What Actually Stops the Payment
Here is the uncomfortable truth behind every big loss: the detection tools can all be right and the money still leaves, because a human receives the request, feels the urgency and acts. The control that actually stopped losses in case after case is not a model, it is a process that verifies the request itself.
- Mandatory out-of-band callback - any payment, banking-detail change or unusual data request must be confirmed through a second, independently established channel, a call to a known number on file, never the number or address in the request. A cloned voice cannot pass a callback to a line the attacker does not control.
- Approval chains that match reality - who can authorize what, up to which amount, with which second signer, written down and enforced, so no single person can move large sums on a single instruction.
- A codeword or shared secret for executives - a simple pre-agreed phrase for high-value verbal requests that no public recording would reveal.
- Friction on urgency - a rule that urgency and secrecy are themselves triggers for more verification, not less, because pressure is the attacker’s main tool.
- Deepfake readiness drills - run the scenario with your finance team the way you run a fire drill, so the first time they see a deepfaked CFO is not during a real attack.
Why process beats detection
In the Arup case, the fraud unravelled only when the employee eventually checked with headquarters3. A mandatory callback before the first transfer would have stopped it on day one. The FBI and most security guidance treat out-of-band verification of payment and banking-detail changes as the single most effective control against business email compromise. Detection narrows the odds. Process stops the loss.
The problem with process is that it lives in people’s heads and in documents nobody reads under pressure. That is precisely where an AI employee that knows your real processes changes the equation.
Where Superkind Fits
Superkind is not a deepfake detector. It is not a media-forensics model, a voice-detection engine or an identity-verification product, and it should not pretend to be. Its role in this space is the one job none of the detection tools do: being the verified-request backstop that knows your company.
- Knows your real approval chains - Superkind’s Company Brain holds who can authorize what, up to which amount, with which second signer, as it actually works in your company, not as a generic rulebook.
- Knows your payment processes - it understands how a legitimate payment or a banking-detail change really flows through your business, so a request that skips a step stands out.
- Catches the mismatch before money moves - when an instruction arrives that does not match your process, the Company Brain flags it, even when the voice on the call sounds perfect, because it is checking the request against reality, not the media.
- Enforces the out-of-band step - it can require the callback, the second signer or the codeword to be completed before a payment proceeds, turning your written policy into an enforced control.
- Connects to your real systems - email, finance, ERP and approval workflows, so it sees the request in context rather than as an isolated message.
- Keeps the institutional knowledge - the fraud reasoning and the “this is not how we do it” instinct stays in the Company Brain when experienced staff leave.
- Builds on your detection tools - it sits on top of whatever media, voice or email detection you run, taking their signals as inputs to a process check, not replacing them.
- Process-first, not model-first - the starting point is always your actual workflows, mapped with your team, which is the only way a backstop can tell a real request from a fake one.
| Layer | What it checks | Example tools | Where Superkind sits |
|---|---|---|---|
| Media detection | Is the audio or video fake? | Reality Defender, Sensity, Intel | Consumes the signal |
| Voice & liveness | Is the caller a live human? | Pindrop, iProov | Consumes the signal |
| Email defense | Is the email impersonation? | Abnormal, Proofpoint, Microsoft | Consumes the signal |
| Process verification | Does the request match our process? | Callback rules, approval controls | This is Superkind’s job |
Superkind
Pros
- ✓ Closes the last gap - verifies the request, not just the media
- ✓ Knows your real processes - the Company Brain, not a generic policy
- ✓ Enforces out-of-band checks - turns written policy into a hard control
- ✓ Layers on your stack - takes detection signals as inputs
- ✓ Retains fraud knowledge - stays when experienced staff leave
Cons
- ✗ Not a deepfake detector - you still need media, voice or email detection
- ✗ Needs process access - only works if it knows how you really operate
- ✗ Not a self-serve app - requires engagement with our team
- ✗ Overkill for the tiny - a manual callback rule may be enough under a handful of staff
How to Choose: Layered Defense by Company Size
You do not need every layer on day one. You need the ones that match where you are actually attacked and what you can run. Here is a practical way to sequence it.
| Your situation | Start with | Add next |
|---|---|---|
| Small company, few payments | Out-of-band callback rule, email defense you already own | Staff drill, codeword for executives |
| Mid-sized, moves money on instruction | Behavioral email defense + enforced approval chains | Process backstop, voice detection if you run a call centre |
| Call centre or customer onboarding | Voice and liveness detection (Pindrop, iProov) | Media detection for high-risk video, process backstop |
| Bank or regulated financial firm | Full stack: media, voice, email and process | Independent evaluation and EU AI Act documentation |
| Public media or platform | Content-scale media detection (Reality Defender, Hive) | Synthetic-media labelling per EU AI Act Article 50 |
Acting Now vs Waiting
Acting Now
- ✓ The cheapest layer is free - a callback rule and a drill cost almost nothing
- ✓ Attack volume is rising - vishing and voice-clone attempts climbed sharply through 2024-20254,5
- ✓ Regulation is arriving - EU AI Act transparency rules on synthetic media are in force19
- ✓ One loss pays for years of defense - a single CEO-fraud event dwarfs the tooling cost
Waiting
- ✗ Your voice is already public - conference talks and voicemails are training data
- ✗ The fakes keep improving - detection-only defenses age fast
- ✗ Staff turnover erases instinct - the “this is not how we do it” knowledge walks out the door
- ✗ You learn under fire - the first drill becomes a real incident
The Deepfake Readiness Playbook
A good defense is a sequence, not a single purchase. Here is the order that gives you the most protection for the least spend, starting with the controls that cost nothing.
- Write the callback rule - make out-of-band verification mandatory for every payment, banking-detail change and unusual data request, to a known number on file. This is the highest-leverage control and it is free.
- Map and enforce approval chains - document who can authorize what, with which second signer, up to which amount, and make the rule binding in your finance workflow, not just on paper.
- Harden email first - turn on and tune the impersonation controls in the email platform you already run, then add a behavioral layer like Abnormal if targeted BEC is getting through.
- Add voice detection where you take calls - if a call centre or finance team authorizes on the phone, layer in Pindrop-style liveness so a cloned voice is scored in real time.
- Add media detection for high-risk video - for executive video approvals or high-value onboarding, add Reality Defender or an equivalent as a signal into the decision.
- Install the process backstop - put an AI employee that knows your real approval chains and payment processes on top, so a request that does not match your process is caught before money moves.
- Run the drill, then run it again - rehearse a deepfaked-CFO scenario with finance and reception, measure who catches it, and fix the gaps. Repeat quarterly.
Deepfake and CEO-Fraud Readiness Checklist
- Out-of-band callback is mandatory for every payment and banking-detail change
- Approval chains are documented and enforced in the finance workflow
- Executives have a pre-agreed codeword for high-value verbal requests
- Urgency and secrecy are treated as triggers for more verification, not less
- Your email platform’s impersonation controls are turned on and tuned
- Voice or liveness detection covers any channel where you authorize by phone
- A process backstop checks requests against how your company really operates
- Finance and reception have run a deepfake drill in the last quarter
A note on GDPR and the EU AI Act
For German and EU buyers, check EU data residency, a signable data processing agreement, and whether a vendor trains shared models on your data. The EU AI Act’s Article 50 requires that AI-generated or manipulated media be marked as such19, and the BSI publishes practical guidance on deepfake threats that is a useful input to a risk assessment22. Lawful synthetic media must be disclosed, so unlabelled synthetic media inside a high-stakes request is itself a signal.
Frequently Asked Questions
There is no single best tool, because deepfake defense is four different jobs. For real-time media detection the specialists are Reality Defender, Sensity AI, Intel FakeCatcher and Hive. For voice and liveness verification, Pindrop leads on audio and iProov on face liveness and injection-attack detection. For business email compromise, Abnormal Security, Proofpoint, Microsoft Defender for Office 365 and Mimecast are the main names. And no detection model replaces an out-of-band process that verifies the request itself before money moves. The right pick depends on where you are actually attacked.
Partly, and not reliably on its own yet. Reality Defender and Intel FakeCatcher can score a live video or audio stream and flag likely synthetic media, and Pindrop scores voice during a call rather than only at the start. But the Arup Hong Kong case showed a full video call of deepfaked colleagues fooling a trained finance employee, and MITRE ATLAS has documented live deepfake injection evading both passive and active liveness checks. Treat live detection as one signal, not a verdict, and back it with an out-of-band verification step for any payment or data request.
Pricing varies by category and is mostly quote-based for enterprise deals. Real-time media detection platforms like Reality Defender are sold as enterprise contracts or per-call API pricing. Voice and liveness tools such as Pindrop and iProov are priced per call, per check or per seat. Email security runs per mailbox: Microsoft Defender for Office 365 is roughly 2 US dollars per user per month for Plan 1 and 5 for Plan 2, while Proofpoint, Abnormal and Mimecast are quote-based. Always model total cost including integration and the staff time to act on alerts.
CEO fraud is a type of business email compromise (BEC) where an attacker impersonates a senior executive to pressure an employee into an urgent, confidential payment or data transfer. BEC is the broader category, which also includes vendor invoice fraud, payroll diversion and account compromise. Deepfakes made CEO fraud far more convincing because the attacker can now add a cloned voice or a video call to the spoofed email. The FBI reported 21,442 BEC complaints and 2.77 billion dollars in losses in the United States in 2024 alone.
It is growing fast. Deloitte projects that generative-AI-enabled fraud losses in the United States could reach about 40 billion dollars by 2027, up from 12.3 billion in 2023. CrowdStrike recorded a 442 percent rise in voice phishing between the first and second half of 2024, and Pindrop reported a sharp spike in deepfake voice fraud attempts. The most-cited single case remains the 2024 Arup incident, where a finance employee in Hong Kong paid out about 25 million dollars after a video call with deepfaked colleagues.
Liveness detection inside an identity verification product covers onboarding and login, but it does not protect a payment-approval conversation or a help-desk reset call. Gartner predicted that by 2026, deepfakes would push 30 percent of enterprises to stop treating face biometrics as reliable in isolation. Treat deepfake defense as layered: liveness and injection-attack detection for identity moments, voice detection for the call centre, media detection for high-risk video, and a process check on top of all of it.
For email, Microsoft Defender for Office 365 fits companies already on Microsoft 365, and both Proofpoint and Mimecast operate EU data regions. For identity and liveness, iProov is widely used in European eID programmes. The main checks for a German Mittelstand buyer are EU data residency, a signable Auftragsverarbeitungsvertrag, whether the vendor trains shared models on your data, and documentation that satisfies the EU AI Act transparency rules on synthetic media. The BSI has published guidance on deepfake threats that is a useful reference for a risk assessment.
Out-of-band verification means confirming a request through a second, independently established channel, like calling a known number on file rather than the number in the email. It matters because every detection tool can be wrong, but a cloned voice cannot pass a callback to a number the attacker does not control. The FBI and most security guidance treat a mandatory callback for any payment or banking-detail change as the single most effective control against CEO fraud. Detection narrows the odds, process stops the loss.
No, and it should not claim to. Superkind is not a media-forensics model, a voice-detection engine or an identity-verification product. It is the AI employee layer that knows your company. Its value in this space is as the verified-request backstop: it holds your real approval chains, payment processes and who-can-authorize-what, so a request that does not match your actual process gets caught before money moves, even when the voice on the call sounds perfect. You still need a detection tool for the media itself.
Not well. The categories are built on different data and different models. Some identity vendors bundle liveness and injection detection, and some email suites add impersonation protection, but no single product leads in media detection, voice detection and BEC email defense at once. Most companies end up layering a detection tool, an email security platform and a verification process. The integration and the process around the tools is where the real protection lives, not in any one engine.
Vendor-claimed accuracy is high but should be read carefully. Intel says FakeCatcher is about 96 percent accurate on its test data, and Pindrop claims 99 percent against known voice-cloning models and 90 percent against new ones. These are vendor figures on their own benchmarks, often without a clear false-positive rate, and accuracy drops against novel attacks the model has not seen. Run a proof of concept against your own calls and media before you trust any headline number, and never let a passing score alone authorize a payment.
Buying a detector and assuming the problem is solved. The tool scores the media, but a human still receives the request, feels the pressure and moves the money. Companies that invest only in detection and skip the process work, the callback rule, the approval-chain enforcement and the staff drills, keep the same exposure with a better dashboard. The most expensive failures, including Arup, were social-engineering wins against people and process, not pure technology failures.
It depends on the layer. An email security platform can be live in days to a few weeks, especially an API-based tool like Abnormal that does not change your mail routing. Voice detection in a call centre and real-time media detection take longer because they need integration and tuning. The out-of-band verification process and staff training can start immediately and cost almost nothing. Start with the process, add email defense next, then layer detection where your highest-risk flows are.
Yes, in part. Article 50 of the EU AI Act sets transparency obligations requiring that AI-generated or manipulated audio, image and video content be marked as artificially generated or manipulated, with limited exceptions. That helps the honest ecosystem but does nothing to stop a criminal who ignores the rule. For a business, the practical takeaway is that lawful AI media must be disclosed, so unmarked synthetic media in a high-stakes request is itself a warning sign worth building into your process.
Related Articles
- The Best AI Tools for Fraud Detection and Financial Crime in 2026
- AI Agent Security: How to Deploy AI Agents Without Opening New Risks
- Prompt Injection: The Security Risk Every AI Deployment Faces
- The AI Email Inbox: How AI Agents Handle Business Mail
- The Approval Bottleneck: Why Decisions Stall and How AI Fixes It
Sources
- FBI Internet Crime Complaint Center (IC3) - 2024 Internet Crime Report
- Deloitte Center for Financial Services - Generative AI and the Risk of Deepfakes in Banking (2024)
- Fortune - Arup Confirmed as Victim of USD 25 Million Hong Kong Deepfake Scam (2024)
- CrowdStrike - 2025 Global Threat Report (vishing up 442%)
- Pindrop - Voice Intelligence and Security Report (deepfake voice fraud)
- iProov - Threat Intelligence Report (injection attacks up 783%)
- iProov - Dynamic Liveness Achieves CEN/TS 18099 High for Injection Attack Detection (2026)
- Intel - FakeCatcher Real-Time Deepfake Detector (via VentureBeat)
- Reality Defender - Multimodal Real-Time Deepfake Detection
- Sensity AI - Deepfake Detection Platform
- MITRE ATLAS - AML.CS0033: Live Deepfake Image Injection to Evade Mobile KYC
- Abnormal Security - Behavioral AI for Business Email Compromise
- Proofpoint - Email Attacks Drive Record Cybercrime Losses in 2024
- Microsoft - Defender for Office 365
- Mimecast - Email Security, Sandboxing and Archiving
- Gartner - Deepfakes Will Undermine Face Biometric Authentication by 2026 (Akif Khan)
- The Register - Gartner on AI Deepfakes and Biometrics (Akif Khan quote, 2024)
- World Economic Forum - Deepfakes and the Erosion of Digital Trust (2025)
- EU AI Act - Article 50: Transparency Obligations for Synthetic Media
- FTC - Voice Cloning Challenge
- Gartner - Deepfake Social Engineering Attacks Survey (2025)
- BSI (Bundesamt fuer Sicherheit in der Informationstechnik) - Deepfakes: Threats and Countermeasures
Ready to close the gap deepfake detectors leave open?
Book a 30-minute call with Henri. We will map where a verified-request backstop fits your approval chains and payment processes - no commitment, no sales pitch.
Book a Demo →
