AI Voice Scams and Deepfakes: The Social Engineering Your Staff Aren't Trained For

Every phishing module written in the last decade taught employees to look for the same things: a misspelt sender domain, clumsy grammar, an odd greeting, a link that does not match its label. Those tells were reliable because the attacker was a stranger typing in a hurry. Generative AI has removed all four of them, and it has added something phishing never had — the voice and face of someone the employee trusts and reports to.

This is not a future risk being briefed to Indian boards as a hypothetical. Voice-cloning fraud is now routine enough that the Reserve Bank of India and SEBI have both publicly cautioned the market about deepfake audio and video circulating in the names of their own senior officials, and the Ministry of Electronics and Information Technology has issued advisories to intermediaries on synthetically generated content. When regulators have to warn people that videos of the regulator are fake, the assumption underneath your security awareness programme — that a familiar voice or face is a form of authentication — has already failed.

What does a deepfake or voice-clone attack look like now?

Three variants account for most of what Indian organisations are seeing. They are usually combined.

The voice clone

A few seconds of usable audio is enough. That audio is rarely stolen; it is published. Conference panels on YouTube, a podcast appearance, an investor call recording, a LinkedIn video, a company culture reel, a voice note forwarded in a WhatsApp group. The clone is then used on a phone call or, more commonly, as a WhatsApp voice note — which is worse for the defender, because a voice note removes the possibility of the employee asking a question mid-sentence and hearing the model stumble.

The deepfake video call

The higher-effort version puts a synthetic face into a live meeting, sometimes with additional fake “colleagues” present to manufacture consensus. Its purpose is almost never to conduct a long conversation. It is to survive ninety seconds, because ninety seconds is all that is needed to convert a suspicious instruction into a verified one in the employee’s mind. The employee later reports, entirely honestly, that they saw the CFO approve it.

The context layer

What makes both of the above land is the accurate detail wrapped around them. Attackers now open with things they should not know: the name of the acquisition target, the fact that the CFO is genuinely in Singapore this week, the vendor code, the internal name of the approval workflow, the fact that the finance controller is on leave. Some of that comes from breached data. Much more of it comes from what the organisation itself publishes — press releases, tender documents, employee posts about travel and reorganisations, and out-of-office replies that helpfully name the person covering.

Why do existing controls and training fail to stop deepfake fraud?

Because they were built to stop error and to help people spot fakes, not to stop a convincing instruction that appears to come from the most senior person in the chain. Four gaps explain why.

Authority short-circuits process

Most Indian finance and procurement controls are designed to stop error and unauthorised action. They are not designed to stop authorised-looking action requested by the most senior person in the chain. When the instruction appears to come from the CFO or the promoter, the junior employee’s realistic choice is between breaching a control and challenging a superior on a Friday evening. Organisations that have never explicitly told staff which one to pick have already made the decision for them.

Urgency and secrecy are treated as facts, not signals

Every one of these frauds contains a deadline and a confidentiality instruction, because both exist to prevent verification. Yet in most awareness training, “urgent” and “do not discuss this with anyone” are presented as characteristics of phishing emails rather than as hard stops that apply regardless of who is asking and how convincing they sound.

The training taught detection instead of verification

This is the central design flaw. Traditional modules ask the learner to spot the fake. Against synthetic media, spotting is an unreliable skill and getting less reliable every quarter — the visual artefacts people are taught to look for, unnatural blinking, mismatched lip movement, flat lighting, are exactly what each new model generation removes. Detection is the wrong objective. The only durable control is out-of-band verification: a defined action the employee performs on a separate channel that the attacker does not control, executed the same way whether or not anything felt suspicious.

Nobody rehearsed the awkward part

A verification rule that exists on paper still fails if the employee has never practised saying the sentence out loud. “I’ll call you back on the number in the directory before I action this” is socially difficult to say to a senior leader for the first time under pressure. It has to be scripted, sanctioned in writing by that leader, and rehearsed.

Which Indian laws apply to deepfake and voice-clone fraud?

There is no single “deepfake law” in India, which leads some teams to conclude the area is unregulated. The obligations are real; they are just distributed across four instruments.

The Information Technology Act, 2000

Under the Information Technology Act, 2000, section 66C covers fraudulent use of another person’s electronic signature, password or unique identification feature — identity theft. Section 66D covers cheating by personation using a computer resource, punishable with imprisonment of up to three years and a fine of up to one lakh rupees. Both apply squarely to voice-clone and deepfake fraud. Section 43A and the wider framework carry the civil consequences of failing to maintain reasonable security practices.

CERT-In directions

The CERT-In Directions of 28 April 2022, issued under section 70B(6) of the IT Act, require organisations to report specified cyber incidents to CERT-In within six hours of noticing them, to maintain logs for 180 days within India, and to enable a point of contact for such reporting. A social engineering fraud that involves unauthorised access, data compromise or a fraudulent transaction can fall inside that reporting scope. Six hours is a materially shorter clock than most Indian incident-response runbooks are actually built to meet.

The DPDP Act, 2023

Where the fraud is achieved by extracting personal data — employee records pulled from an HR system through a convincing call to a helpdesk, for instance — the organisation is dealing with a personal data breach as a Data Fiduciary under the Digital Personal Data Protection Act, 2023, with notification duties to the Data Protection Board and to affected Data Principals, and no materiality threshold to hide behind. Section 8(6) of the Act covers every personal data breach. Rule 7 of the DPDP Rules, 2025, which applies eighteen months after the Rules were notified on 13 November 2025, requires intimation without delay and a detailed report to the Board within 72 hours. We covered the training side of that obligation in our analysis of why the DPDP Act’s biggest gap is employee-level awareness.

Sectoral and payment obligations

Regulated entities carry additional duties: RBI’s cyber security and digital payment security directions, customer-protection and fraud-reporting circulars, and turnaround-time expectations. For any organisation, the practical first step after a fraudulent transfer is immediate escalation to the bank and reporting through the National Cyber Crime Reporting Portal or the 1930 helpline. Recovery, where it happens at all, happens inside the first few hours.

Your employees were trained to detect a fake. The attacker’s entire investment now goes into being undetectable. Train the callback instead.

Which controls stop deepfake and social engineering fraud?

Five controls do most of the work: a callback rule that removes judgement, written leadership backing for pushback, hardened high-risk transactions, simulations beyond email, and fast, blameless reporting. All five can be implemented this quarter.

1. Write a verification rule that removes judgement

One sentence, published, applying to everyone: any instruction to move funds, change bank details, share credentials or release data that arrives by call, voice note, video call or message is verified by calling the requester back on the number held in the internal directory — never a number supplied in the request — before it is actioned, irrespective of seniority or stated urgency. Add a value threshold above which a second named approver must independently confirm. The point of writing it this way is that it requires no assessment of authenticity, which is precisely the assessment employees can no longer reliably make.

2. Have leadership pre-authorise the pushback in writing

The CFO, CEO and promoter should state, on the record and in their own words, that they expect to be called back and will never treat verification as insubordination or as a lack of trust. This single message does more for control effectiveness than any module, because it removes the social cost that causes the control to be skipped.

3. Harden the specific transactions attackers target

Vendor bank-detail changes, one-off “confidential” transfers, payroll account changes, gift-card and prepaid purchases, and credential or MFA resets at the service desk. Each needs a channel-independent confirmation step: a callback to a previously verified vendor contact, a documented second approver, or in-person or manager-attested identity confirmation for helpdesk resets. Attackers pick these because they are the points where a single person’s decision moves money or access.

4. Simulate beyond email

If your phishing programme only sends emails, it is testing the one channel the attacker has largely left behind. Extend simulations to voice calls, voice notes, WhatsApp and meeting invitations, and measure the behaviour that matters — the percentage of recipients who performed the callback and reported the attempt — rather than only click rates. Role-relevant information security awareness training that covers synthetic-media social engineering, with a short annual information security refresher to keep the verification habit current, gives you both the coverage evidence and the per-person completion record an auditor or a customer questionnaire will ask for. Our information security training hub sets out how these courses fit together.

5. Make reporting fast and blameless, then drill it

One number, one mailbox, reachable outside working hours, with an explicit no-blame commitment for anyone who reports having been deceived. Then run a timed tabletop: a cloned-voice payment instruction is actioned at 5.15 pm on a Friday and discovered at 6.00 pm. Who calls the bank, who files on the cyber crime portal, who assesses the CERT-In six-hour trigger, who decides whether the DPDP Act notification duties are engaged, and who tells the board? An organisation that has never timed that sequence will discover its answer during a real incident, when the recoverable window has already closed. The same logic we applied to why a strong firewall does not compensate for untrained employees holds here, only with a shorter clock.

How can you test whether your verification control is real?

Pick five people who can move money or grant access — in finance, procurement, payroll and the IT service desk. Ask each one a single question: if the CFO calls you right now, sounds exactly like himself, tells you this is confidential and asks you to release a payment before the cut-off, what do you do, and what exactly do you say?

If any of the five describes trying to work out whether the voice is real, the programme is still training detection and the control does not exist. If all five describe the same callback, name the number they would use, and can say the sentence out loud without hesitating, the control is real — and the clone on the other end of the line stops being a fraud and becomes an incident report.