TL;DR: Deepfake-enabled fraud is exploiting the gap between system authentication and human-to-human trust, with Arup’s Hong Kong office losing about $25 million in one incident and forecasts putting annual losses at $40 billion by 2027, according to SlashID. Perception-based verification is no longer enough; identity programmes must account for cryptographic proof in live communication as well as login events.
At a glance
What this is: This analysis argues that deepfake impersonation is breaking perception-based trust in IAM and that live, human-to-human verification needs cryptographic proof, not voice or face recognition.
Why it matters: IAM teams, IGA leads, and security architects need to treat impersonation over calls and video as an identity control problem, because login MFA does not verify who is on the other end of a live conversation.
Context
Deepfake impersonation is a human identity problem that sits outside the protection envelope of conventional login controls. If a fraudster can convincingly imitate a colleague in a live call or video meeting, the normal IAM assumption that authentication happens at the system boundary is no longer enough.
The article focuses on the gap between perception-based trust and cryptographic proof. That matters for human IAM, because many high-risk decisions still happen in conversations, not just in applications, and those conversations can now be manipulated with synthetic voice and video.
Arup’s Hong Kong fraud is a useful example of why this is not an abstract future risk. The article treats that case as representative of a growing class of impersonation attacks, not an isolated anomaly.
Key questions
A: Voice recognition and caller ID break down because both are easy to spoof or synthesize. A cloned voice can sound convincing enough to pass casual judgment, and caller ID can be faked cheaply. When teams trust either signal, they create a path for attackers to reset credentials, transfer MFA, or approve access without proving identity.
Q: Why do deepfakes create fraud risk even when MFA is already deployed?
A: MFA protects system access, not human trust inside a live conversation. A fraudster can pass the login boundary and still impersonate a colleague on a call or video meeting. That is why organisations need a second assurance layer for communication-time identity, especially where approvals or transfers can happen outside applications.
Q: What are the signs that conversational identity controls are failing?
A: Common warning signs include urgent payment requests, unusual escalation paths, staff accepting requests because the voice sounds familiar, and approvals made without any independent proof. If the organisation has no audit trail for who verified whom during the conversation, the control is effectively absent.
Q: How should security teams handle identity verification in high-risk video calls?
A: Security teams should treat high-risk video calls as identity checkpoints, not just collaboration sessions. Use participant verification before approving hiring, account recovery, payments, or privileged changes. The goal is to confirm the human behind the screen and the integrity of the camera source before any trust-sensitive decision is made.
Technical breakdown
Why login MFA does not solve live impersonation
Multi-factor authentication proves a user’s identity to a system, but it does not prove that the person speaking on a phone or video call is genuine. Deepfake attacks exploit that separation by shifting the trust decision away from the login screen and into a social interaction where visual and auditory cues are easy to falsify. The result is a control gap between system authentication and conversational trust. In IAM terms, the identity event is still real, but the decision to release information or approve an action happens in a different channel with no equivalent assurance level.
Practical implication: treat high-risk conversations as an identity surface and define a separate verification control for them.
How bidirectional TOTP changes human verification
Mutual TOTP applies the familiar TOTP concept in both directions so each participant proves possession of a registered device during the conversation. Instead of trusting a voice or face, both parties generate time-bound codes from unique secrets and verify them live. That creates a short-lived trust exchange with device-bound proof, rather than relying on perception or pre-existing familiarity. The important architectural shift is not the code format itself, but the fact that verification is mutual and occurs during the interaction where the fraud attempt would otherwise succeed.
Practical implication: use bidirectional verification for sensitive conversations where one-sided identity checks are not enough.
Why the trust boundary must move from perception to cryptography
Deepfake impersonation works because humans are taught to trust continuity of voice, face, context, and conversational fluency. Cryptographic verification breaks that assumption by making trust dependent on possession of a registered secret and device, not on how convincing the person sounds. That is why real-time impersonation becomes much harder to scale when the communication channel demands proof rather than recognition. For identity architects, the lesson is that the trust boundary now has to extend into the meeting itself, not stop at application authentication or endpoint login.
Practical implication: redesign trust decisions so they depend on proof of possession, not subjective human recognition.
Threat narrative
Attacker objective: The attacker wants the victim to trust a synthetic colleague long enough to disclose sensitive information or approve high-value action.
- Entry occurs through phishing or social engineering that brings the target into a seemingly legitimate live conversation.
- Credential access is replaced by identity mimicry, where AI-generated voice and video provide the attacker with a convincing human persona.
- Escalation happens when the target shares sensitive information or authorises financial action inside the trusted interaction.
- Impact is fraudulent transfer or privileged disclosure that occurs before traditional security controls can intervene.
Breaches seen in the wild
- Arup deepfake fraud 2024: Deepfakes of Arup's CFO and colleagues on a video call led a Hong Kong employee to transfer HK$200 million (about US$25.6m) to fraudsters.
Read our 52 NHI Breaches Analysis report for a comprehensive view of breaches impacting Non-Human Identities including AI Agents.
NHI Mgmt Group analysis
Perception-based trust is no longer a viable identity control for high-risk conversations. The article shows that deepfakes do not need to break authentication systems to succeed; they only need to hijack the human trust layer that sits around them. That means IAM programmes must stop treating voice, face, and familiarity as sufficient evidence of identity. The practitioner conclusion is simple: live communication needs its own assurance model.
Cryptographic proof must enter the human interaction layer. Traditional identity design assumes proof is needed at login, then trust can be carried forward into the session or conversation. Deepfake fraud breaks that assumption because the risky action happens after login, during the call itself. The implication is that identity governance has to cover communication-time verification, not only access-time verification.
Mutual verification is the right pattern when both parties are exposed to impersonation risk. A one-way challenge still leaves one side vulnerable to a convincing fake on the line. Bidirectional proof aligns better with sensitive conversations because both participants confirm each other before the interaction becomes consequential. For practitioners, the governance question is no longer whether a caller is known, but whether both sides can prove they are legitimate at the moment it matters.
Human identity controls need to be designed for deception, not just authentication failure. The article’s core message is that fraudsters are exploiting the social assumptions embedded in enterprise workflows, especially around finance and privileged decisions. That creates a named control gap: conversation trust debt. The longer programmes rely on subjective recognition, the more exposure accumulates in places MFA cannot see. The practitioner conclusion is to retire perception as a control signal for sensitive actions.
Deepfake impersonation is pushing identity security toward cryptographic assurance outside the application boundary. This is not just a fraud problem or a training problem. It is a sign that identity governance must extend across channels, because attackers now operate inside the trusted human layer. Teams that only harden login flows will miss the place where the compromise actually occurs. The conclusion is that communication verification belongs in the identity model.
What this signals
Conversation verification is becoming part of identity governance. For IAM teams, the important shift is that identity assurance can no longer end at login. If an organisation approves payment changes, privileged requests, or recovery actions in live calls, those interactions need a proof step of their own.
Conversation trust debt: the more often staff rely on recognition instead of verification, the more opportunity deepfake impersonation has to succeed. That debt accumulates in finance, help desk, and executive support workflows, where urgency and familiarity often suppress challenge.
Programmes that already invest in phishing resistance, privileged access controls, and human identity assurance should now extend that thinking into live communications. The control question is not only who logged in, but who was verified before a consequential conversation turned into an action.
For practitioners
- Define a high-risk conversation policy Classify finance approvals, password resets, privileged requests, and vendor payment changes as interactions that require explicit identity verification before any action is taken.
- Deploy mutual verification for sensitive calls Require both participants to prove device possession before sharing sensitive information in meetings, voice calls, or ad hoc escalation channels.
- Separate recognition from authorization Stop allowing familiarity with a voice or face to satisfy approval conditions, especially where money movement or access changes are involved.
- Add impersonation scenarios to fraud playbooks Test how staff respond when a synthetic executive, supplier, or administrator joins a live conversation and asks for urgent action.
- Log communication-time verification events Record who verified whom, when the proof occurred, and whether the conversation was cleared for sensitive action so investigators can reconstruct the trust decision.
Key takeaways
- Deepfake impersonation shifts the problem from system authentication to human trust, which means login controls alone cannot protect sensitive conversations.
- The article uses the Arup fraud to show that synthetic voice and video can drive real financial loss before traditional security tools have a chance to intervene.
- The most direct control change is to replace perception with cryptographic proof when a conversation can trigger money movement, privileged access, or sensitive disclosure.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and MITRE ATT&CK address the attack and risk surface, while NIST SP 800-63, NIST CSF 2.0 and OWASP ASVS set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI09 — Human-Agent Trust Exploitation | Deepfake impersonation exploits human trust during high-stakes interactions. |
| Recommendation — Treat synthetic impersonation as a trust-exploitation risk and add explicit verification to sensitive conversations. | ||
| MITRE ATT&CK | TA0001;TA0006 — Initial Access; Credential Access | The article maps the attack to impersonation and preparation techniques. |
| Recommendation — Map deepfake-enabled social engineering to impersonation and credential-access behaviours in detection content. | ||
| NIST SP 800-63 | SP 800-63C — Federation | The article deals with identity proofing across a communication relationship, not just a login event. |
| Recommendation — Extend federation assurance thinking to any identity assertion that drives a high-risk decision. | ||
| NIST CSF 2.0 | PR.AA-05 — Access Permissions, Entitlements and Authorizations | The article is about deciding who may trigger sensitive actions after identity verification. |
| Recommendation — Tie high-risk actions to explicit authorisation rules and require stronger proof before approval. | ||
| OWASP ASVS | V10 — OAuth and OIDC | The article centres on identity assurance mechanics, though at the human communication layer. |
| Recommendation — Use higher-assurance identity flows where sensitive decisions depend on verified identity assertions. | ||
Key terms
- Deepfake-based impersonation: A fraud technique that uses synthetic audio, video, or both to make an attacker appear to be a trusted person during a live interaction. The tactic exploits human trust in familiar cues and often aims to trigger urgent actions such as payments, resets, or access changes before verification is challenged.
- Mutual Verification: A two-way identity check in which both parties prove who they are before a sensitive interaction continues. In practice, this means the requestor and the approver both rely on cryptographic or out-of-band proof rather than visual or auditory cues alone.
- Perception-Based Trust: A trust model that relies on human cues such as voice, face, familiarity, or conversational style to decide whether someone is genuine. It is fragile in modern fraud scenarios because those cues can be imitated convincingly, while the underlying identity proof remains unverified.
- Communication-Time Identity: The identity assurance that applies during a live conversation rather than only at system login. It matters because many sensitive decisions happen in calls, chats, and meetings after authentication has already occurred, leaving a gap that impersonation attacks can exploit.
Deepen your knowledge
NHI governance, human identity, and identity lifecycle management are core topics in our NHI Foundation Level course, the industry's only accredited NHI security programme. If you are building or maturing an IAM programme, it is worth exploring.
Published by the NHIMG editorial team on August 15, 2026.
Updated on October 7, 2026.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org