Verified Person, Unverified Authority: AI and the Next Gap in Digital Trust

In August 2026, Singapore’s police and cybersecurity agency warned of a cryptocurrency-related scam that had already caused US$11.8 million in losses. The attack began not with an ingenious exploit of computer code, but with something more mundane: professional trust.

Someone impersonating a recruiter from a cryptocurrency company approached a victim on LinkedIn. Emails followed from a domain closely resembling the legitimate firm's, along with several video interviews. Eventually, the victim was asked to complete a coding assessment, and he performed it on a company-issued device. The assessment was hosted on a spoofed website, malicious software was installed, internal credentials were compromised, and cryptocurrency was transferred out.

The sequence is instructive. Recruiters approach candidates online. Interviews take place by video. Developers are given coding assessments. Each apparently ordinary interaction lent credibility to the next. This was not simply a failure of identity verification. It exposed a more fundamental weakness in the architecture of digital trust.

Four different questions

Much of the investment in digital identity has concentrated on answering one question: who are you?

That matters. Governments are particularly well placed to establish high-assurance identity. Singapore’s Singpass and MyInfo ecosystem, for example, allows individuals to authenticate themselves and, with consent, provide participating organisations with information drawn from government sources.

But many commercially significant interactions depend on four distinct propositions:

• Identity: Who are you?
• Affiliation: Whom do you represent?
• Authority: What are you entitled to do on their behalf?
• Intent: Did you actually approve this particular action?

The distinctions are consequential. Conclusive proof that John Smith really is John Smith says nothing about whether he works for the bank listed on his online profile. Proof that he works there does not establish that he is authorised to recruit employees, request confidential information or instruct a transfer of funds. And even an appropriately authorised officer may not have approved the transaction now presented in his name.

Digital systems have become increasingly sophisticated at establishing the first proposition while frequently relying on much weaker evidence for the other three. Professional platforms illustrate both the weakness and the beginnings of a solution. LinkedIn offers workplace verification through mechanisms including corporate email addresses and Microsoft Entra Verified ID. The more important development is conceptual: proving that a person exists and proving that the person has a particular institutional affiliation are separate tasks.

The same distinction becomes more important as artificial intelligence changes the economics of deception.

When credibility becomes cheap

Fraud does not require perfect impersonation. It requires impersonation sufficiently convincing to persuade enough people at sufficiently low cost.

Historically, sophisticated deception was expensive to manufacture. Fraudulent profiles had thin employment histories. Phishing emails contained grammatical mistakes. Conversations were generic. Supporting material took time to fabricate. None of these signals was foolproof, but collectively they raised the cost of sustaining a credible false identity.

Generative AI is eroding that advantage. Professional biographies, personalised approaches, convincing correspondence, synthetic images and prolonged conversations can all be produced more cheaply and repeatedly. The FBI recorded almost US$21 billion in reported internet-crime losses in 2025, including more than US$17.7 billion associated with cyber-enabled fraud. The shift matters because much of this fraud attacks trust rather than software. Business email compromise, investment fraud and impersonation succeed when victims accept false signals of identity, affiliation or authority.

In the Singapore case, one warning sign was that the supposed interviewer kept his camera switched off during the interviews. It would be unwise to treat that as a durable defence. The FBI has already warned of scams involving synthetic images, voice cloning and believable fake videos. What is an obvious anomaly today may be considerably less obvious tomorrow.

Fake recruitment operations aren't confined to opportunistic scammers. Microsoft has documented the “Contagious Interview” campaign, in which attackers construct realistic recruitment processes before inducing developers to run malicious packages or commands. Fireblocks has identified related recruiter-impersonation tradecraft associated with North Korean actors. Singapore authorities have not attributed the August attack to those actors, and doing so would be wrong. The significance lies in the technique: recruitment itself has become an attack surface.

This creates an asymmetry. The cost of manufacturing credibility is falling, while the burden of assessing it still rests heavily on human judgement.

The answer cannot simply be to train everybody to become a better lie detector.

From asserted identity to verifiable claims

A more robust architecture would assign different claims to institutions capable of making them authoritatively.

Governments and other high-assurance identity providers can establish personal identity. Employers can attest that an individual currently works for them. Organisations can establish that person’s role and scope of authority. For sufficiently consequential actions, the individual can separately authenticate the transaction itself. Parts of this architecture already exist. The Global Legal Entity Identifier Foundation’s verifiable Legal Entity Identifier, or vLEI, is designed to enable digital verification of the relationship between legal entities and the people acting for them, including organisational roles. ISO 17442-3 standardised the underlying framework in 2024.

The broader direction is more important than any particular technology. Digital verification is beginning to move from exchanging copies of documents and accepting assertions towards exchanging claims that can be independently verified. Yet stronger verification creates an obvious privacy problem. A system that responds to impersonation by encouraging every platform to accumulate passports, national identity numbers, employment files and organisational records would exchange one vulnerability for another.

The better objective is to prove more while revealing less.

Selective disclosure allows a person to reveal only the attributes relevant to a particular interaction. Zero-knowledge proofs can go further, allowing a proposition to be proved without revealing all of the information underlying it. The European Digital Identity Wallet framework makes selective disclosure a central design feature, while its evolving technical architecture is also exploring how to incorporate zero-knowledge proofs as the technology develops.

The distinction matters. A counterparty may need to establish that somebody has a valid high-assurance identity, currently holds a credential issued by Company X and is authorised to perform a specified function. It does not follow that the counterparty requires that person’s identification number, residential address, complete employment record or other unrelated information.

Zero-knowledge proofs will not be necessary everywhere. In many situations, an ordinary digitally signed credential with selective disclosure will be simpler and entirely adequate. Privacy-enhancing technology should not become an excuse for technological extravagance. Nor will implementation be frictionless. A functioning credential ecosystem needs trusted issuers, interoperable standards, accurate organisational records and reliable mechanisms for checking status. Companies must determine who can issue credentials and who can revoke them. Systems have to communicate across organisational and national boundaries. Smaller firms may struggle to justify the infrastructure.

These are serious practical constraints. They are not a convincing argument for relying on self-assertion indefinitely.

The neglected problem of time

Even a perfectly authentic credential can become dangerously misleading when it becomes stale.

An employee leaves. A director changes role. Signing authority is withdrawn. A recruiter’s mandate expires. A legitimate account is compromised. The integrity of a credential therefore depends not only on how securely it was issued, but on whether it can expire or be revoked reliably. A claim that was completely accurate six months ago can be more dangerous than an obviously questionable one precisely because it carries greater credibility.

Verification without effective revocation risks producing little more than a highly convincing stale credential.

This also explains why cryptographic credentials cannot replace conventional controls. A genuine CFO can still be deceived. An authorised employee’s computer can still be compromised. A properly verified recruiter can have an account hijacked. High-consequence actions will continue to require segregation of duties, transaction monitoring, independent confirmation and other forms of defence in depth.

The objective is not to manufacture certainty. It is to stop asking one weak signal to perform several different jobs.

When does good practice become expected practice?

This raises a question that legal, risk and compliance functions should begin considering. New controls do not become regulatory requirements simply because the technology exists. Cost, proportionality, interoperability and the severity of the underlying risk all matter. But expectations of reasonable protection do not remain static either.

Singapore’s Shared Responsibility Framework offers an instructive analogy. It does not require verifying professional affiliation or authority. Instead, it allocates specified responsibilities among financial institutions and telecommunications companies for a defined category of phishing scams. The broader principle is nevertheless significant: fraud prevention is increasingly treated as a problem involving multiple participants in a trust ecosystem, rather than one whose consequences can simply be left with the person who was deceived.

Organisational identity may eventually raise the same question. If reliable affiliation and authority credentials become cheap, interoperable and widely available, at what point does relying solely on self-declared employment information cease to be reasonable for high-risk interactions?

There is no universal regulatory answer today. Firms need not wait for one before examining where their own trust architecture is weakest.

A practical review can follow the same four questions:

• Identity: where does the organisation need high-assurance proof of who an external party actually is?
• Affiliation: where can a claimed relationship with an employer, adviser, vendor or counterparty create access or credibility?
• Authority: which consequential actions presently depend on assumed rather than independently established authority?
• Intent: for which transactions should the authorised person have to confirm the particular act through an independent channel?

A fifth question cuts across all four: how quickly do credentials, permissions and trusted relationships expire when circumstances change?

Technical controls should follow the risk. Microsoft, for example, recommends running coding assessments from external parties in dedicated, isolated environments rather than on corporate devices containing sensitive credentials or privileged sessions. Elsewhere, a secondary call-back, authoritative corporate directory or independently verified instruction may be sufficient.

The control should fit the proposition being tested.

From plausible to provable

The lesson from the Singapore attack is therefore larger than “be careful with recruiters on LinkedIn”. The fraud worked because credibility accumulated across a sequence of plausible interactions. A professional profile supported an affiliation. The affiliation legitimised an email. The email lent credibility to the interviews. The interviews made the coding assessment appear routine.

The attacker climbed the trust stack one assertion at a time. As AI makes those assertions cheaper to manufacture, digital trust will have to rely less on how convincingly a claim is presented and more on whether the appropriate party can verify it.

Government-backed identity can help establish who someone is. Employers can establish whom they represent. Organisations can establish what they are authorised to do. Authentication can help establish whether they approved a particular act. Privacy-preserving technologies may increasingly allow these propositions to be tested without requiring everyone involved to surrender ever more personal information.

None of this can prove that a verified person is honest, prudent or exercising good judgement. Technology can authenticate an actor and an authority; it cannot authenticate wisdom. That limitation matters. But before deciding whether to trust what somebody is asking us to do, we should at least be able to establish who is asking, whom they represent, and whether they have the authority to ask.

In an age of synthetic credibility, the future of trust may depend less on making the plausible easier to recognise than on making the important claims easier to prove.

This personal opinion piece by Clemence Kng, co-authored by Shann Ng, is exclusive on RegTech.com.