In February of last year, the CEO of a mid-market logistics firm took a Teams call from the chair of his board. Two members of the audit committee were on the screen beside her. The chair explained a time-sensitive acquisition opportunity — confidential until close, legal was drafting the paperwork — and asked the CEO to authorize a wire through finance before the European markets opened. The faces on the screen were faces he had sat across from for three years. The voices were the voices he had heard on a dozen calls the month before. Both audit-committee members nodded when the chair asked if they agreed with the timing.
The CEO approved the wire. $25 million. Gone in fifteen transactions over the next four hours.
None of the people on that call were real. Every face, every voice, every nod had been generated — trained on years of public video and earnings-call audio, rendered in real time, scripted by criminals who had studied the firm's internal culture for weeks. There was no malware. No system was breached. The network logs showed a CEO authorizing a legitimate wire from his legitimate workstation, fully authenticated, in a conference call with three directors he believed were his directors.
The technology worked exactly as designed. The laptop worked. The video codec worked. The bank's fraud controls worked. Every system on the chain did its job.
What failed was belief.
The Wrong Conversation
For most of the last three years, the conversation about deepfakes has been a technology conversation. Can the video be detected? Can the audio be flagged? Is the lighting off? Does the mouth sync? Are there artifacts in the ears, the fingers, the shadows?
Those were useful questions in 2023. In 2026 they are a trap. The generation side is iterating faster than the detection side and always will. Every tell you can train someone to spot today will be ironed out by a model release next month. We have been here before with phishing emails — the era of "look for typos" turned out to be a narrow window that closed the moment language models learned to write English.
The deepfake-detection arms race is not a fight you are going to win on the viewer's side of the screen. Any defense that depends on an executive squinting at a CFO's mouth and deciding whether the sync looks right is a defense that was obsolete before it was deployed.
Which means the interesting question is not Can we spot the deepfake? It is Why did anyone act on it?
The Real Attack Surface
Deepfakes do not break systems. They break trust calibration.
Every organization runs on a quiet assumption: when you see a familiar face and hear a familiar voice, you are talking to the person that face and voice belong to. That assumption has been load-bearing for the entire history of business. It is what lets the finance team move money at the speed of email, what lets executives delegate without paperwork, what lets companies function without everyone independently verifying everything.
For two centuries that assumption was safe, because impersonating a face and a voice in real time was functionally impossible. Synthetic media killed that guarantee in about eighteen months. The assumption is still there, baked into every approval workflow and every culture of "the boss said so, just get it done." The attackers noticed before most companies did.
The attack surface here is not your video conferencing platform. It is the moment where a CEO sees his board chair's face and his brain marks the request as verified. That moment happens inside the human, not inside the machine. No endpoint agent, no zero-trust architecture, no deepfake-detection filter sits between the face and the decision.
Why Leadership Is the Softest Target
It would be convenient if the people most at risk were the easily duped. They are not. The people most at risk are the senior and senior-adjacent employees who have been trained, correctly, to move fast and trust the system.
A CEO's assistant who second-guesses every request from the CEO is a bad assistant. A finance director who refuses to process a confidential wire until she has personally chased down every party to the deal is a bottleneck. A board chair who has to prove his identity on every call is an insult. Organizations explicitly teach their senior people to take cues from authority, to prioritize velocity, to extend trust inside the circle. Deepfakes are designed to exploit exactly this conditioning.
This is why the old security-awareness framing misses the executive floor almost entirely. Awareness campaigns have been aimed at the lower-risk, higher-volume population — the employees clicking links and opening attachments. That is where the attacks used to land. The attacks have moved. The highest-dollar fraud of the last two years has run through the C-suite and their direct reports, not the help desk.
The executives running these organizations were never trained to doubt a familiar face. They were trained to trust their teams. Those are not the same thing, and the gap between them is where a $25 million wire lives.
What Trust Calibration Actually Looks Like
Calibrating trust is not the same as being paranoid. It is not seventeen approvals on every request. It is a small number of disciplined behaviors, practiced often enough that they survive the moment a high-stakes request goes live.
- Out-of-band verification for anything that moves money, access, or people. If a familiar face on a video call asks for a wire, the receiver calls that person back on a pre-established channel before anything moves. Not a number suggested on the call. Not a new Teams thread. A channel that existed before the request. Every time. The friction is the point.
- A published map of who can request what, through which channel. If a wire above a threshold can only be authorized through one specific workflow, a deepfake asking for it outside that workflow has nowhere to land. The rules need to be explicit and visible inside the company, not an assumption in one executive's head.
- A shared language for "I want to slow this down." Employees — including senior ones — need a low-cost way to say "let me confirm this separately" without it reading as insubordination or mistrust. If that phrase gets someone scolded once, the organization has just told everyone else not to use it.
- Leaders saying the quiet part out loud. Executive teams that openly discuss "what would a deepfake of me ask you to do differently than I would" give their people a reference frame for catching the thing when it happens. Silence on the topic gives them nothing.
None of this requires new software. None of it requires detecting the deepfake. It requires a culture where the natural response to an unusual, high-stakes, authority-driven request is not do it fast but verify first, even when the face on the screen is familiar.
Why Annual Training Will Not Get You There
Executives are the hardest group to move with traditional training. They skip the modules. They delegate the compliance checkbox. They have, understandably, less patience than anyone else in the building for a 45-minute video on the generic threat landscape.
But calibrated trust is not knowledge — it is habit. Habits come from repetition, and repetition means short, frequent, pointed reps, not an annual compliance event. Five minutes a week on a realistic scenario does more to rewire the reflex than a day-long offsite on the threat landscape. This is the same argument behind our cyber risk briefings for leadership teams and Sentinel Weekly for the broader organization: the goal is not to describe the risk. It is to install the reflex.
By the time the deepfake call comes in, knowledge-level awareness is too slow. The only thing fast enough is the trained instinct that says I'm not sure, let me verify before the conscious mind has caught up with why.
The Point
Deepfakes are not the problem. They are the delivery vehicle. The problem is that most organizations still run on unexamined trust — trust in a face, trust in a voice, trust in the momentum of a conversation already in progress — and that trust has not been recalibrated for a world where all three can be synthesized in real time.
You cannot patch belief. You cannot install a filter on the part of the brain that marks a familiar face as safe. What you can do is build a small number of verification habits into the organization and rehearse them until they fire automatically — especially under pressure, especially from authority, especially when the face on the screen is a face your people have known for years.
If you want an honest read on where your own organization stands, the Sentinel Vault Cyber Checkup is five minutes and it does not flatter anyone. It will not tell you whether your people can spot a deepfake. Nothing will. It will tell you whether your people have been given permission to doubt one.
The attack surface is not your network. It is the moment someone familiar asks for something unusual, and somebody else decides whether to believe them.