The technology behind synthetic media has followed a familiar trajectory: years of academic research, a period of gradual improvement, and then a threshold crossing that made previously theoretical attacks operationally viable for adversaries without specialised technical capability. That threshold crossed somewhere between 2023 and 2024. By 2025, the tools required to clone a voice from a short audio sample, generate real-time video deepfakes, and produce synthetic documents convincing enough to pass casual inspection had become accessible through consumer applications and open-source repositories.

The $25.6 Million Problem

The Hong Kong finance case is worth examining in detail because it illustrates the specific failure mode that deepfakes introduce. The employee received an invitation to a video call that appeared legitimate. The participants on the call included what appeared to be the CFO and other colleagues known to the employee. The conversation followed a plausible narrative: an urgent, confidential business transaction required a transfer. The employee, confronted with a video call showing familiar faces and voices giving explicit instructions, executed the transfer.

Every control the organisation had built for this type of request failed. The request came through a legitimate-looking communication channel. It was accompanied by visual and audio confirmation that appeared to come from senior leadership. The urgency framing that typically triggers suspicion in a text-based phishing email reads differently when it is delivered in what appears to be a live conversation with the CFO. The callback procedure that might have caught a text-based fraud was rendered irrelevant because the employee believed they were already in a live call with the authorising executive.

The finance worker was not negligent. They followed the visual and social evidence in front of them. The problem is that the evidence was fake and there was no control in place that could distinguish real from synthetic at the moment of the transaction.

Why Existing Controls Are Structurally Insufficient

The standard enterprise response to social engineering has three layers: technical controls at the communication channel level, such as email filtering and caller ID verification; procedural controls at the transaction level, such as dual authorisation and callback procedures; and awareness training that teaches employees to recognise suspicious signals. All three layers were designed for an adversary communicating in text or voice, not one generating real-time synthetic video.

Technical controls have no native capability to detect AI-generated video in real time. Email filters analyse headers, content, and sender reputation. They cannot analyse whether the face in a Teams call is real. Voice authentication systems trained on biometric voice profiles may be defeated by voice cloning from a small audio sample. The controls that exist are auditing the wrong layer.

Procedural controls assume that verification means reaching the actual person. A callback procedure is effective against a fraudster pretending to be the CFO via email because calling the CFO directly reveals the deception. It is ineffective when the fraudster can answer the callback with a cloned voice. Dual authorisation procedures assume that two independent people cannot both be deceived simultaneously. In a coordinated deepfake attack where both appear to be present on the same video call, that assumption breaks.

Awareness training faces the hardest problem. It can teach employees to be sceptical of urgent requests, unfamiliar senders, and pressure tactics. It cannot teach an employee to detect a real-time deepfake through observation alone, because the artefacts that distinguish synthetic video from real video in 2026 are below the threshold of human perception in most viewing conditions. Training that tells employees to watch for unnatural blinking or edge distortions is providing outdated guidance for a technology that has already passed those detection thresholds.

What the 2026 Threat Landscape Requires

The most honest framing of the deepfake problem is that it requires controls at a layer that most organisations have not yet built: authentication of content origin rather than authentication of sender identity. Verifying that a communication comes from a real person in real time, rather than assuming it does because the channel is trusted, is a fundamentally different security problem from the ones the industry has spent two decades building solutions for.

Several technical approaches are in development. Cryptographic content provenance, including specifications under the Coalition for Content Provenance and Authenticity (C2PA) standard, embeds verifiable metadata about the origin and chain of custody of digital content. If widely adopted, this would allow a video call platform to indicate whether the video feed carries a verified provenance signature or does not. The limitation is adoption: C2PA requires participation from device manufacturers, platform providers, and content creation tools simultaneously to be effective.

Liveness detection and deepfake detection APIs are available from several vendors, but they are probabilistic rather than definitive, they require integration into communication platforms that were not built with this use case in mind, and their accuracy degrades as the underlying generation models improve. This is a recurring problem with detection-based approaches to AI-generated content: the detection capability is always chasing a generation capability that is moving faster.

The most operationally reliable response in 2026 is process-level: redesigning high-stakes financial and access-granting workflows so that they cannot be authorised through a single communication channel, regardless of how convincing that channel appears. Out-of-band verification through a pre-established, independently confirmed contact method, combined with time delay requirements for high-value transactions, does not defeat deepfakes technically but removes the operational window in which they succeed.

What Security Leaders Should Do Now

• Audit which financial and access approval workflows can be completed through a single communication channel. Any workflow that can be authorised through a video call, voice message, or email alone, without an independent out-of-band confirmation, represents an exposure to deepfake attack.
• Establish pre-registered, independently confirmed verification contacts for executives and high-privilege individuals. The callback procedure needs to route to a number registered in the organisation’s directory independently of the communication that initiated the request.
• Review voice and video authentication systems. Any system that uses biometric voice or video as a primary authentication factor should be assessed against the current capability of voice cloning and real-time video synthesis tools.
• Include deepfake scenarios in security awareness training. Employees should understand that visual and audio evidence in a video call is not a reliable indicator of participant identity in 2026. The training objective is not detection but scepticism: teaching employees to apply process controls regardless of how confident they feel about who they are talking to.
• Track C2PA and content provenance standards adoption in the communication platforms the organisation uses. Platforms that implement content provenance verification will provide a technical layer that does not currently exist. Procurement decisions should weight this capability.