The Security Risks of AI Clones: How Executive Digital Doubles Threaten Corporate Safety

The technology is advancing faster than the systems used to establish whether the person on a screen is actually the person who appears to be there. Video calls, familiar voices and executive presence have traditionally functioned as informal layers of trust inside organisations. Synthetic media is turning all three into potentially unreliable signals.

The end of video as proof of presence

The $25 million fraud at engineering company Arup is the clearest warning in the briefing. An employee joined a video conference populated by deepfake versions of senior executives and was persuaded to transfer the money to Hong Kong accounts.

The significance of the incident lies less in the sophistication of the fake than in the failure of the organisation's assumptions. The conference looked like a normal corporate interaction. The people appeared to be familiar colleagues. Yet the visual and auditory evidence that would once have made the interaction credible had become part of the attack.

That changes the security problem fundamentally.

Traditional corporate controls were designed around the idea that credentials can be protected and communications authenticated. A password can be replaced. A compromised account can be locked. A suspicious transaction can trigger a second approval.

A convincing synthetic executive creates a different vulnerability: it attacks the human decision-making layer itself.

The scale of the threat is already visible in fraud data. Deepfake verification company Onfido has reported a 3,000 per cent increase in deepfake scams over recent years. The Arup case shows what happens when that technology crosses from identity fraud into a high-value corporate decision.

The practical question for companies is no longer simply whether an employee has received a legitimate message from an executive. It is whether an organisation can prove that a real human authorised the instruction in the first place.

## Delegating identity before controlling it

The technology behind these systems is not mysterious. A personal clone can combine a large language model with private documents and communications, instructions about how its owner communicates, memory systems and voice or video generation. Some advanced platforms can create a usable avatar from as little as 15 seconds to two minutes of recorded material.

That low barrier to creation is one reason adoption is accelerating.

But the same architecture introduces several points of failure. A clone may expose information from indexed private material. Its underlying model may hallucinate. An injected instruction may alter its behaviour. And once the system is allowed to operate without a human checking every interaction, there is no guarantee that an apparently authoritative statement was ever approved by the person whose identity it represents.

HeyGen has already encountered a version of this problem: an executive clone overpromised enterprise plans while the company's founder was on paternity leave. CarynAI, an audio clone created from influencer Caryn Marjorie's identity, bypassed its intended boundaries and engaged in explicit interactions before being terminated.

Neither incident is comparable to a multimillion-dollar corporate fraud. But together they expose the same structural weakness. The software can reproduce the appearance and style of a person without reliably reproducing that person's judgement.

That distinction becomes critical when the clone is no longer merely answering questions but negotiating, approving, selling or directing.

An AI assistant helps a worker perform a task. An agentic clone begins to represent the worker in the task.

The difference is not cosmetic. It is a transfer of authority.

Identity is becoming an asset

The most consequential shift may therefore be legal rather than technical.

Once a voice, face, mannerisms and accumulated professional knowledge can be converted into software, a person's identity becomes something that can be licensed, commercialised and deployed continuously. That creates obvious opportunities for people with valuable public profiles. It also creates a new class of rights disputes for ordinary workers whose bargaining power may be far weaker.

The response from high-profile individuals is already moving towards formal ownership. Matthew McConaughey trademarked his voice and likeness in January 2026, followed by Taylor Swift in April.

For AI developers, such moves matter because they push the question of synthetic identity into intellectual property law. A model may be technically capable of reproducing a person's voice or appearance. Whether it should be allowed to do so, under what conditions and for whose benefit is a separate question.

Celebrities have the resources to establish legal boundaries around their likenesses. Most employees do not.

That gap is becoming more important as companies experiment with AI clauses and the licensing of employee identities. Workers can be asked to provide recordings, communications, and workflow data that make their own replacement easier to automate. Startups such as Action Model are already offering white-collar workers as much as $100 an hour to capture task data and workflow steps for AI agents.

The economic bargain is uncomfortable: a worker may be paid to teach a system how to perform the job, while helping create the conditions under which the human role becomes less valuable.

Entertainment has already exposed the same tension, with backlash over synthetic performers such as Tilly Norwood and growing disputes over AI clauses. What begins in Hollywood or modelling can become a much broader workplace issue once digital likenesses become ordinary enterprise assets.

Security will have to move beyond passwords

The immediate institutional problem is that most organisations still treat identity as something that can be recognised rather than cryptographically established.

That model becomes increasingly fragile when an attacker can reproduce the face and voice of a senior executive convincingly enough to conduct a live conversation. A video call can no longer serve as meaningful proof of presence merely because participants can see and hear one another.

Companies will need stronger ways of establishing provenance and authority around high-risk communications. The briefing points towards cryptographic identity verification and real-time biometric provenance tracking as the direction of travel.

But technology alone will not resolve the underlying question.

A verified avatar could prove that a system is authorised to represent a person. It would not necessarily prove that a particular statement was personally approved by that person. The distinction becomes especially important when a clone is allowed to operate autonomously.

If an AI representative promises a customer a contract term that does not exist, responsibility cannot simply disappear into the model. The commercial relationship still has a human or corporate owner, and the legal consequences of the machine's statement have to land somewhere.

That creates an unresolved liability problem for the platforms and companies deploying these systems.

The same technology that allows one executive to appear simultaneously in hundreds of conversations also makes it possible to manufacture hundreds of apparently authoritative statements without the executive ever knowing they were made.

The economic case for delegation will remain powerful. A digital double can extend a person's reach at a fraction of the cost of physical presence.