TL;DR
- In a 2025 Gartner survey of 302 cybersecurity leaders, 62 per cent said their organisation had experienced at least one deepfake attack in the prior 12 months.
- 43 per cent reported a deepfake audio call incident and 37 per cent encountered one on a video call.
- The FBI warned in May 2025 that cloned voices can sound nearly identical to the real person and advised confirming identity via an independently sourced number.
- A January 2026 Osterman Research study found 60 per cent of security and IT leaders lacked confidence in their ability to counter deepfake attacks even with existing training.
- A structured drill, run on your own schedule with nothing real at stake, is how you find out whether your people actually follow verification rules under pressure.
The call comes in at 4:40 on a Friday. It sounds exactly like the CFO.
She is travelling, rushed, apologising for the bad connection. She needs a supplier payment pushed before the weekend. Ten minutes later an email lands with the wire details. For a brokerage, the same shape of attack looks slightly different: a call asking to change bank details on a settlement, or to reset a lender portal login. If the person who can authorise that action is also the person who answers the phone, there is no second desk to catch it.
Is hearing the fake a realistic defence?
Security awareness training has long taught staff to spot tells: awkward phrasing, an odd sender address, a link that does not match the text. Voice clones leave fewer of those tells every quarter.
The FBI made this point directly in its May 2025 warning about AI-generated voice messages impersonating senior US officials. Cloned voices can sound nearly identical to the real person. The FBI told recipients to confirm identity by calling an independently sourced phone number before acting on anything, not the number supplied in the request.
The January 2026 Osterman Research study of 128 US security and IT leaders found 60 per cent lacked confidence in their ability to counter deepfake attacks even with the training they already run. That is not a training problem. That is a process problem. If an employee runs the same verification step every time money, credentials, or permissions are involved, the quality of the deepfake stops mattering much.
This is the same principle behind Agent Validation: Stop Before You Book, Charge, or Send: irreversible actions need a confirmation gate, whether the instruction comes from a human voice or an automated system.
What does a drill actually involve?
The source article from The Hacker News sets out a practical structure. Start with two short lists.
The first is the people an attacker would impersonate: executives who can approve payments or direct staff, then one level down to VPs and directors in finance, IT, and HR. Add outside contacts with real authority over money or access. The second list is the people an attacker would call: accounts payable, treasury, payroll, HR, executive assistants, and the IT help desk. The help desk gets left off these lists often. It should not be. A cloned voice asking for an MFA reset is a fast route to account takeover.
For everyone on the first list, note how much public audio and video exists. The article notes it does not take hours of audio. A few minutes from a single webinar can be enough, which puts most directors on the list.
Before running any scenario, document the verification rules employees are supposed to follow. A drill without a clear rule only measures instinct. Strong rules share a few traits: out-of-band callback using a number from the internal directory, clear thresholds for which actions always require verification, dual approval for new payees or changed banking details, and explicit written permission from the CEO and CFO for staff to hang up and call back without fear of consequences. Deepfake fraud runs on urgency and hierarchy. An employee who thinks slowing down a VP will cost them at review time is the easiest target in the building.
What scenarios should you test?
Real attacks rarely arrive as a single phone call. They usually start in email or chat to set context, then move to voice or video to apply pressure. Drill scenarios should follow the same path.
The article outlines four starter scenarios: an urgent wire from the CFO, a supplier bank change confirmed by a follow-up call, a help desk MFA reset from a travelling VP, and a new admin account request ahead of an audit. Before building any of these, get sign-off from legal, HR, and every executive whose identity will be used.
The output of a drill is not a score. It is a list of the exact moments where the verification step did not happen, and the conditions that made it easy to skip. Fix those conditions. Then run it again.
For brokers already thinking about how AI systems handle sensitive instructions, the post on SalesBleed: What a Hidden Prompt in an Enquiry Form Did to Salesforce's AI Agent shows what happens when an automated system acts on an instruction without a confirmation gate. The failure mode is structurally identical to a deepfake call: a convincing instruction, no verification step, an irreversible action.
The advice from the source article is worth keeping: do not just do something, stand there. A minute of standing still and one callback costs nothing. Acting on the deepfake costs considerably more.
Originally published at theautomate.io.
Top comments (0)