Beyond Detecting Fakes: Assuring Identity And Verifying Authorization
Krishnaveni Palanivelu, SVP Cybersecurity Architect, Financial Services.
gettyImagine a bank’s fraud team reviewing two incidents. In the first, an applicant cleared onboarding using a synthetic identity assembled from stolen personal details and fabricated supporting documents. In the second, a finance employee released a payment after hearing an AI-cloned version of a familiar executive’s voice. The executive never requested it.
Both involved convincing fakes. But they exposed different failures: accepting an identity that did not represent a real applicant and accepting instructions that did not come from the person supposedly giving them. Stronger onboarding addresses the first. It cannot, by itself, prevent the second.
Generative AI reduces the time, effort and expertise required to create believable documents, faces, voices and messages. The FBI has warned that criminals use these capabilities to expand fraud and make their approaches more convincing. The Financial Crimes Enforcement Network has also reported suspicious activity involving fraudulent identity documents used to bypass verification.
Attackers can vary their material, repeat attempts and exploit whichever channel offers the weakest challenge. Training employees to notice an unnatural pause or a distorted image may help today, but it is a fragile foundation for tomorrow’s payment controls.
Detection still matters. However, making it the deciding control leaves the organization dependent on recognizing the next forgery before accepting it. Leaders should ask: What independent evidence must exist before we establish a relationship or allow a consequential action?
Identity assurance establishes confidence that someone is who they claim to be. It connects the applicant, supporting evidence and subsequent credentials to a real identity. The next checks concern authorization and intent: Is this person permitted to act, and what evidence shows they knowingly approved this specific request?
A customer can authenticate successfully and still be manipulated into paying a criminal. An employee can recognize a voice and mistake familiarity for permission. Authentication establishes control of a credential, but it does not establish informed intent.
In my article on non-human identity, I argued for attributable actions, defined permissions and accountable ownership across machines and agents. The same discipline must govern human interactions: establish who is acting, preserve that identity through the workflow and verify authority at the point of consequence.
In banking, account opening and authorized-push-payment fraud require different interventions. Onboarding should corroborate identity evidence against trustworthy sources and examine whether the applicant, device and submitted material belong together. A payment from an established customer requires transaction context: a new beneficiary, an unusual amount or signs that someone is directing the customer through the process. Another login challenge may add little protection against persuasion.
In healthcare, identity assurance can help prevent an impostor from gaining access to a patient portal, but releasing records or changing a caregiver’s access requires a separate authorization decision. Confirming the patient’s identity does not automatically establish that another person should receive their information.
Benefits administration presents a similar split. Establishing that a claimant exists does not prove eligibility, and verifying eligibility does not authorize every later change. Redirecting a benefit payment should trigger verification tied to that change, using contact details established before the request.
Across these settings, assurance belongs at the decision where harm becomes possible.
I would start with the highest-consequence journeys: opening an account, recovering access, changing payment details, releasing sensitive records and moving money. For each, identify the evidence required, the accountable owner and the conditions that should pause execution.
Layer identity checks deliberately. Document validation, authoritative-source checks, liveness, device signals and behavioral history should contribute different evidence. Five checks derived from the same compromised document do not create five independent reasons for confidence. Liveness detection, which confirms a real, live person is present (rather than a photo, replay or synthetic video), also needs protection against injected media; a live-looking face is insufficient if the capture process itself can be bypassed, for example, by injecting a pre-recorded or AI-generated video directly into the system rather than presenting it to a real camera. NIST’s identity-proofing guidance explicitly addresses forged media and injection attacks.
Consequential actions require confirmation through a separately established channel. A voice or video call should never, by itself, authorize a payment or sensitive change. Present the actual recipient, amount and action in a trusted application, and require approval bound to those details.
Independence is essential. Calling a number supplied during the suspicious conversation simply returns the organization to the attacker. Account recovery and contact detail changes deserve equivalent scrutiny because they can undermine every later confirmation.
Even a transaction-specific approval cannot prove that someone understands a scam. Unusual payments may require a cooling-off period, a targeted warning or a conversation with a trained specialist.
I previously argued that security reviews should follow information through its lifecycle. Fraud assurance needs that perspective too.
Trace where identity evidence originated, who changed contact details, which device initiated the request and how the destination differs from previous activity. Behavioral signals should inform investigation and proportionate challenges, with privacy limits and a route for legitimate users to appeal mistakes.
Govern your own generative AI with equal care. Inventory customer-facing assistants and internal agents, restrict their access and prevent conversational output from becoming authority to change records or release funds. Test whether manipulated inputs can persuade an assistant to bypass verification or expose information useful for impersonation.
These controls carry costs. Additional checks can delay customers, exclude people with limited documentation and generate false positives. Provide accessible alternatives and measure abandonment and mistaken rejection alongside fraud outcomes.
As I argued in my ORF America commentary, cybersecurity affects economic security because failures undermine the services and confidence on which commerce depends.
The fraud metrics reported to the board should lead with realized fraud losses reported both before recoveries and after recoveries, measured against transaction value and broken out by fraud type. Track customer harm, recovery time and legitimate activity blocked alongside them. Detection rates remain diagnostic measures; they are not the outcome. The practical test is whether fewer fraudulent actions succeed, less money is lost and legitimate people can still get things done.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

