AI and Cyber Crime: What Changed in 2025–26
AI cyber crime in 2025–26: how voice cloning, deepfake video calls, AI-written phishing and scam compounds changed fraud in India, and what still works.

In short
- AI did not invent new scams. It removed the three things that used to give scams away: bad grammar, a stranger's voice and a stranger's face.
- Voice cloning now needs a few seconds of audio, which most people have published. Treat an urgent voice as unverified until you call back on a known number.
- Deepfake video calls are used in digital arrest and corporate payment fraud. A live "officer" or "director" on video proves nothing.
- The industrial side matters more than the clever side: scam compounds run scripts, translation and targeting at scale, and launder through stablecoins.
- The defences that survive are procedural, not technical: a call-back rule, a family safe word, a second approver, and the 1930 reflex.
Who this is for: Anyone trying to understand why scam calls, messages and videos suddenly got more convincing, including families, employees, investigators and the people who write policy for all three.
For twenty years, three things gave an online scam away: the grammar was wrong, the voice belonged to a stranger, and the face, if there was one, was a stock photo. Between 2024 and 2026, generative AI removed all three. That, more than any new trick, is the AI cyber crime story: the old frauds, with their tells sanded off, at a volume no call centre could manage before.
This explainer is the hub for our emerging-threats coverage. It sets out what actually changed, what did not, what Indian regulators have said, and which defences survived. Each section links to a deeper piece.
What AI cyber crime looks like in practice
Start with the scams Indians already know, then add one AI component to each.
| Familiar scam | What AI added | Read more |
|---|---|---|
| "Family emergency" call | The caller's voice is your son's, cloned from a reel | Voice clone scam |
| Digital arrest | A live "CBI officer" on a deepfake video call, in a studio set, with a generated face | Digital arrest explained |
| Investment tips group | A deepfake video of a known investor or minister endorsing the app | Deepfake investment ads |
| Phishing email or SMS | Perfect Hindi, Tamil or English, in the bank's tone, personalised from leaked data | Phishing in India |
| Business email compromise | A voice or video call from the "director" approving the transfer | BEC explained |
| Sextortion | A face morphed onto explicit video; no recording of the victim was ever needed | Sextortion: what to do |
None of the rows is a new crime. Each is cheating by personation under the IT Act and the BNS, exactly as before. What changed is the production cost, and therefore the scale.
The three mechanisms, explained plainly
Voice cloning
Modern speech models learn the characteristics of a voice (pitch, pacing, the way vowels are shaped) from a short sample and then synthesise new speech in that voice from typed text. A few seconds of clean audio is enough for a clone that passes on a phone line, where compression hides the imperfections. The sample comes from wherever you have spoken in public: an Instagram reel, a YouTube talk, a voice note forwarded in a group, a recorded customer-care call. The clone can speak Hindi even if the original only ever posted in English. The practical defence, a family safe word and a call-back rule, is covered in its own post.
Deepfake video
Face-swap models map one face onto another in real time, well enough for a video call at the resolution WhatsApp allows. The output has artefacts: lighting on the face that does not match the room, glasses that flicker, teeth that blur, a jawline that lags the audio when the speaker turns. Each year the artefacts shrink. The scam compounds that run digital arrest and investment frauds use generated faces for "officers" and "analysts" precisely because a live human on video was the last thing victims trusted. How deepfakes are detected explains the forensic signs and why detection keeps losing ground.
Large language models for scripts and targeting
The quietest change is the biggest. Large language models write AI-written phishing messages with no errors, in any Indian language, in the register of a bank, a courier or a court. They translate a working script from one region to another overnight. They draft replies in a romance scam conversation so that one operator can run dozens of victims at once. And fed with leaked personal data, they personalise: your name, your bank, your last transaction, your child's school. The data that makes that possible is the subject of dark web data markets.
The industrial side: scam compounds and money movement
The clever technology sits on top of an industry. UNODC's regional office for Southeast Asia has documented compounds in Myanmar, Cambodia and Laos where trafficked workers, many from South Asia, run scripted frauds under coercion, with targeting, translation and chat handled by shared tooling. Indian nationals have been among those rescued. Europol's threat assessments describe the same professionalisation in Europe: crime-as-a-service, where scripts, cloned voices, fake apps and money-laundering are bought as modules.
AI slots into that supply chain as one more module. Synthetic identity fraud, where a borrower or account holder is assembled from real and fabricated details, feeds the supply of mule accounts the compounds depend on. It did not create the compounds; it made each worker in them several times more productive. The money then leaves India through mule accounts and is converted to stablecoins, the subject of how crypto and stablecoins launder scam money. The UNODC and Europol EC3 pages describe what each body publishes.
What has not changed
Three things, and they are the basis of every defence that works.
First, the scam still needs you to act. Every AI cyber crime case we have examined ends with a human entering an OTP, approving a transfer, installing an app or sending a document. The technology makes the request more convincing; it cannot make the transfer itself.
Second, the money still moves through Indian bank accounts, which means the 1930 helpline and the CFCFRMS freeze still apply, and the first hour still matters. The reporting sequence in how to report cyber crime is unchanged.
Third, the law already covers it. Cheating by personation using a computer resource (IT Act section 66D), identity theft (66C), publishing obscene or sexually explicit material (67, 67A), and cheating and extortion under the BNS do not care whether the voice was real. The AI and deepfake scams page lists the sections.
What the regulators have said
The MeitY deepfake advisory of December 2023, followed by a further advisory in March 2024, reminded platforms of their obligations under the IT Rules 2021 to act against deepfakes, impersonation and misinformation, and to remove reported content within the prescribed timelines. It has since proposed rules requiring synthetically generated content to be labelled. The advisories are published on meity.gov.in; our MeitY page summarises its role. The content takedown process explains how to use those rules when a deepfake of you or your company appears.
CERT-In has issued advisories on AI-enabled phishing and deepfake-based social engineering for organisations, and I4C's public campaigns have added voice-clone and deepfake examples to the digital-arrest warnings. None of this replaces the household rule above; it supports it.
Where this is going
Expect three things over the next two years. Real-time voice clones that answer questions in conversation, not just read a script. Generated "proof" documents: FIR copies, court orders, bank letters, with correct formatting and seals. And more automation of the victim relationship, so that one operator manages more conversations with less attention, which will make scams both more frequent and slightly more brittle when a victim goes off-script. Detection tools will improve too, but they will keep trailing the generators. The defence does not change: delay, verify on a known channel, and involve a second person.
Next step: tonight, agree a safe word with your family and tell your parents that no officer, bank or relative will ever need money or an OTP during a call. For a workplace, add a second approver on every payment. If a call is in progress right now and someone claims you are under arrest, hang up and read how to escape a digital arrest; if you need help during or after one, our digital arrest response service is built for exactly that.
Frequently asked questions
What exactly did AI change about cyber crime in India?
Mostly cost and polish. Writing a flawless Hindi or English phishing message, cloning a relative's voice, producing a fake video of a well-known investor, or translating a script for a new region used to need skilled people. Now it needs a subscription. The result is more attempts, each more convincing, aimed at more people, rather than a new category of crime.
How much audio does a voice clone need?
Current tools produce a usable clone from a few seconds of clear speech. A WhatsApp voice note, an Instagram reel, a conference talk or a customer-care call recording is enough. The clone can then say anything in that voice, including in a language the original speaker does not use.
Can I tell a deepfake video call from a real one?
Sometimes, and less often each year. Signs include odd lighting on the face, glasses or earrings that flicker, a jaw that does not match the audio, and a refusal to turn sideways or wave a hand in front of the face. Do not rely on spotting it. Rely on a rule: no money or data moves on the strength of a video call alone.
Has the Indian government done anything about deepfakes?
MeitY issued advisories to intermediaries in December 2023 and March 2024 reminding platforms of their duty to act against deepfakes and misinformation under the IT Rules, and has proposed rules on labelling synthetically generated content. Impersonation, cheating and obscene content are already offences under the IT Act and BNS; the deepfake is the method, not a gap in the law.
Is AI also helping the defenders?
Yes. Banks use models to flag mule-account behaviour, platforms use them to detect cloned voices and generated faces, and investigators use them to sift call-data and chat records. But detection tools lag generation tools, and a defence that depends on spotting fakery is weaker than one that does not need to.
If you want help with this
- Digital arrest response · A caller claiming to be police, CBI, ED, customs or TRAI kept you on video and made you transfer money. We stop the pressure, file the right complaints and handle the bank and the follow-up calls.
- Phishing simulation and employee security awareness · Realistic phishing, vishing and WhatsApp simulations with short training that changes what people do, measured over time, and built around the scams actually hitting Indian companies.
- Employee security awareness program (annual, with phishing simulation) · training · An annual programme for all staff with a half-day kickoff, quarterly phishing simulations, short refreshers and a report your auditor and board can read, built around behaviour, not slides.
- Cyber crime investigation fundamentals · training · A three-day course for police, legal professionals and compliance teams on investigating financial cyber crime in India, from the first complaint to money trail, platform requests, evidence and charge sheet.
Keep reading
How Deepfakes Are Detected, and Why Detection Is Losing
Detectors look for the seams: lighting that does not match, lips that lag consonants, metadata that was rewritten. The seams get smaller eve...
AI Voice Cloning Scam: The "Family Emergency" Call
It sounds exactly like your son: an accident, a police station, money needed now. The voice was made from ten seconds of a reel. How it work...
Voice Clone Scams and the Family Safe Word
A cloned voice can say anything. It cannot say the word your family agreed at dinner last Sunday. Here is how to choose one, who should know...
Crypto Money Laundering: How Stablecoins Move Scam Money
The money you lost to a digital arrest or a trading app was probably a stablecoin within a day. Here is the route it takes out of the Indian...