Your CFO on a Video Call: Deepfakes and the New Verification Rules

The first thing to understand about deepfake fraud is that the technology is not the hard part any more. A convincing voice clone can be produced from a few seconds of audio, and the much-quoted McAfee finding that three seconds was enough to produce an eighty-five percent match remains a fair description of what commodity tools can do. Video is harder, but not hard enough to matter when the target is a finance team on a deadline.

The second thing is that this has become an ordinary crime with an ordinary cost. The FBI's 2025 Internet Crime Report, published in April 2026, recorded 22,364 complaints involving AI, with adjusted losses of 893.3 million dollars. Business email compromise with an AI component accounted for more than 30 million dollars on its own. Those figures undercount, because most victims never file and many never realise what they were looking at.

Why It Works

A deepfake attack is not a technology problem; it is an authority and urgency problem that happens to use technology. The caller knows the name of the person they are impersonating, the name of the person they are calling, the deal that is closing and the reason the payment cannot wait. That context is available in the average inbox, and it is what makes the request plausible enough to act on before the scepticism arrives.

Gartner's survey of organisations put the share that had experienced at least one deepfake attempt in the previous twelve months at 62%, covering social engineering, impersonation on a voice or video call, and attacks on automated verification. Surveys of this kind tend to oversample large enterprises and the worried, so treat the precise number as indicative. The pattern it describes is consistent with incident reporting: the calls are short, they target people with payment authority, and they are timed around travel or quarter end.

Voice Is Not an Authenticator

Voice biometrics occupy an awkward position in this. Consumer voice authentication, the kind that lets a customer reset a PIN by speaking, is a low-assurance factor that should not be protecting anything valuable, and replay and synthesis attacks against it are well documented. Enterprise identity verification is a different discipline: document capture, biometric comparison against the document, and liveness detection that has been tested against a defined attack set. The distinction matters because the two get conflated in exactly the wrong direction, with organisations deciding that voice is insecure in general and therefore that nothing should be verified at all.

What can be said with confidence is that hearing a familiar voice, or seeing a familiar face on a video call, is no longer evidence of identity. Any process that treats it as evidence is a process an attacker can rent for a few dollars.

The Law Changed in August 2026

The EU AI Act's transparency obligations under Article 50 applied from 2 August 2026, and the Digital Omnibus regulation that amended the Act in July kept them intact. Providers of generative systems must mark synthetic output in a machine-readable way, and deployers of deepfakes must disclose that the content is artificially generated. Chatbots must make clear that the user is talking to a machine. For a security team the practical consequence is not that fraudsters will comply. It is that the honest half of the market will start emitting provenance data, which gives verification a technical hook it did not have before.

Content provenance standards such as C2PA are worth supporting for the same reason. A signed record of capture is a stronger signal than a detector's probability score, and it fails in a predictable direction: absence of provenance is not proof of forgery, but presence of valid provenance is meaningful evidence.

Controls That Hold Up

• Call back on a number you already have. Not the number on the invoice, not the number in the email, and not the number the caller offers. Use the directory entry or the previous contact record.
• Pre-agree a challenge. A phrase or question established in person, for executives and finance staff, that is not derivable from social media or a breach.
• Separate the approval from the request. Above a defined threshold, payment requires two approvers who verify through independent channels. A single call can never authorise a transfer.
• Slow the urgent path down. The attacker's main asset is time pressure, so make the exception process slower than the standard one rather than faster.
• Do not rely on detection tools. Deepfake detectors are probabilistic, degrade under compression, and are being trained against by the same market. Use them as a signal, never as the decision.

There is also an incident playbook, and most organisations do not have one: what an employee should do mid-call when something feels wrong, how to preserve the recording and the number, who to tell, and how to reach the person being impersonated without using the channel the attacker controls. Rehearse it in the same tabletop exercise as ransomware, because the first ten minutes look identical: an urgent request, an executive unreachable, and someone who wants to help.

The deepfake is not the vulnerability. The vulnerability is a payment process that accepted a voice as proof and a single person as authority, and that was exploitable by a convincing phone call long before the voice could be synthesised.