Can detection software solve this?

Not on its own, and relying on it inverts the problem: it makes the absence of a detection alert into a reason to proceed. Detection is an arms race against generation, and the verification control is not, which is why the control is the durable part.

Is a code word between colleagues a real control?

It is a real control and a fragile one. It works because the attacker does not have it; it fails when it is shared over the channel being attacked, when somebody forgets it under pressure, or when the person under pressure is new. It is a reasonable supplement to a callback and a poor replacement for one.

What if the request really is urgent?

Then it survives a five-minute callback. That is the argument to have once, in advance, with whoever would be annoyed by the delay, because having it during the incident is what the attack is counting on.

What is a deepfake payment scam?

A deepfake payment scam is an attempt to make somebody move money or grant access by impersonating a person they trust, using synthesized voice, video or writing. The mechanism is old and the impersonation is what changed: the signals people used to rely on, such as clumsy writing or a voice that did not sound right, have stopped being reliable. The defense that still works is not better detection. It is confirming the request through a channel the request could not touch.

What actually happens

A request arrives that is plausible, urgent, and from somebody senior enough that questioning it feels like an error. It asks for something ordinary in itself: change these bank details, approve this payment, add this person, reset this access.

The impersonation carries the plausibility. It may be a voice on a call, a face in a meeting, or writing that matches how the person actually writes, assembled from material that is public because that person does their job in public.

The urgency carries the rest. These attacks are built to make the checking step feel like the obstacle, and the person doing the checking feel like the problem.

Why the old tells stopped working

As of 2026, the reliable signal used to be language: awkward phrasing, wrong register, errors a native speaker would not make. That signal is no longer reliable, and nothing has replaced it at the level of the message.

As of 2026, voice and video cloning is good enough, from short public samples, that we would not advise anybody to rely on recognizing a person as a control. We are not going to say how short or how good: the honest answer is that it moves, and the control below does not depend on the answer.

There is also a second-order effect that gets missed. Where AI drafts a company's outgoing mail, its own genuine messages read more like generated text than they used to, which erodes the comparison people were quietly relying on.

We are not going to put a number on how often this happens. The figures we have seen in circulation come mostly from companies selling detection and are not comparable with each other, and the decision does not turn on the number: the control below is worth having at any incidence above zero.

The control that does not depend on noticing

Confirm the request through a different channel, using contact details you already hold rather than any supplied in the request. Calling the number in the message confirms the attacker. This is the whole of it, and it works regardless of how good the impersonation was, which is why it is the control worth having rather than training people to spot fakes.

Give it a second person for anything consequential, and make sure that person is able to refuse. A second approver who cannot say no is a keystroke rather than a control, and whether the seniority runs the wrong way decides which of the two you have.

Then remove the exception. Every control of this kind has an escape path for urgency and seniority, and that path is what the attack is built to trigger. A rule that has been tested on somebody senior and held is worth more than one that is written down.

What to do afterwards, and before

Make reporting a near miss costless. These attempts are commonly made against several people before one works, and the reports from the people who spotted it are usually the earliest warning available. Embarrassment is what stops them arriving.

Practice rather than brief. Reading about a technique produces recognition of that technique; practicing produces the pause, which is what generalizes to the variant nobody has written about yet.

And include the people who are actually targeted. Finance, executive assistants and whoever can grant access are a small group, easy to leave out of a session scheduled for everybody, and the only group the attack was ever going to arrive at.

Common questions

Can detection software solve this?: Not on its own, and relying on it inverts the problem: it makes the absence of a detection alert into a reason to proceed. Detection is an arms race against generation, and the verification control is not, which is why the control is the durable part.
Is a code word between colleagues a real control?: It is a real control and a fragile one. It works because the attacker does not have it; it fails when it is shared over the channel being attacked, when somebody forgets it under pressure, or when the person under pressure is new. It is a reasonable supplement to a callback and a poor replacement for one.
What if the request really is urgent?: Then it survives a five-minute callback. That is the argument to have once, in advance, with whoever would be annoyed by the delay, because having it during the incident is what the attack is counting on.

Find out where you stand

The assessment on this site asks up to 24 questions about what the AI in use here exposes you to, and returns a score, a verdict and, where a control is short, the one thing to fix first. It never asks whether you have a list of the tools. It takes about ten minutes, and questions your organization has no surface for are dropped rather than scored against you.

It is a self-assessment of what has been arranged and practiced, not of whether it would hold against somebody actually trying. Not a penetration test, not an audit, not a certification.

Run the assessment