A familiar story: Your phone rings. The number is real. The voice on the other end is calm and professional, identified as a representative of your bank's fraud prevention department. It is about a $487.32 charge that hit your account at a retail store two states over. You didn’t authorize this, did you? Press one to speak with a fraud specialist.
Four things are happening in that call. Urgency. Specificity. An appeal to a trusted authority. And a call to action that promises to give you control back by signaling that doing nothing puts you at risk.
Every successful scam in 2026 borrows from the same playbook, with new variations every quarter as attackers retool. By the time the call connects, the attack is mostly complete. The work that mattered happened weeks or months earlier, in places most fraud teams aren't looking.
This is the shift reshaping fraud defense, and it is the one most financial institutions are still adjusting to: Fraud has stopped breaking through firewalls and has started breaking down people. Traditional controls, such as firewalls, code scanners, antivirus software, and perimeter defenses remain necessary. However, they are also increasingly insufficient for this particular threat.
That threat actually divides into two problems, and they are worth separating because they fail in different ways.
In the first, an intruder gains access to the account – the classic account takeover – and the question is whether the person in the session is actually the customer. In the second, no one breaks in at all: the legitimate customer authenticates cleanly and is then talked into moving their own money. The call that opens this piece is the second kind.
The two diverge on detection: One looks like anomalous access, the other like anomalous intent from a genuine user, and on who absorbs the loss, since a payment the customer was deceived into authorizing has historically fallen outside the protections that cover an unauthorized transfer, a line that is now beginning to shift. This makes catching the second kind an economic question, not only a service one.
What they share is a single, harder conclusion: A clean login is no longer proof of safety. Authentication has become a checkpoint, not a verdict.
The numbers tell the story plainly. The FBI's IC3 put total reported losses at roughly $20.9 billion in 2025; about 85% of it, or $17.6 billion, was cyber-enabled fraud carried out over digital channels. Account takeover reports in FinCEN Suspicious Activity Reports (SARs) have tripled, according to Dr. David Maimon of Georgia State University. Every dollar of loss costs financial institutions about $5.75 to absorb (as detailed in the LexisNexis True Cost of Fraud Study 2025), including investigation, recovery, regulatory response, and reputational impact.
Only about one in five financial institutions (Fis) reports having mostly or fully automated processes to fight fraud. The rest still lean on human analysts to keep pace with attackers who increasingly are not human at all. In the near term, the automated-attack curve is bending sharply upward: HUMAN Security's 2026 benchmark recorded post-login account-takeover attempts more than quadrupling year-over-year, with automated scraping now behind nearly one in five site visits.
The defenders’ workload is about to be multiplied by a factor that no analyst headcount plan can absorb. Something has to change in the model.
The most useful framework for thinking about modern fraud is the FS-ISAC Cyber Fraud Prevention Framework. It breaks the attack chain into five phases: Reconnaissance (victim identification, target research, infrastructure setup), Initial Access (the latching-on moment when an attacker establishes a foothold), Positioning (changing account information, redirecting communications to controlled channels), Execution (the actual transformation of stolen data into money), and Monetization (the loss realization itself).
Lumin Digital CRO Sean McElroy: Continuous detection.
Most fraud teams have excellent visibility into Phases 3 through 5. Their tooling, analytics, and detection models all live downstream: where transactions move, where money flows, where loss is realized. What they often cannot see, except in retrospect, are Phases 1 and 2. The reconnaissance. The initial access. The earlier signals.
But the earlier signals exist. They are already being generated by the digital banking channel itself, by the cybersecurity tools watching the perimeter, and by the consumer behaviors that subtly shift before an attack escalates.
The opportunity is not in collecting more signals. The opportunity lies in connecting those already being produced.
Some signals are available before an attacker ever touches your login page; call these the no-contact indicators. DNS (Domain Name System) reconnaissance is the earliest of them, though it's noisy and low quality on its own. Data breaches are a much stronger signal (credential exposure that puts your users at elevated risk), but their shelf life is indeterminate.
Breached identity can be weaponized years after the breach itself; you cannot stay vigilant forever on an unbounded timeline. Behavioral signals matter too: Dormant accounts that suddenly become interesting, such as ones with associated liens like HELOCs, can indicate that someone other than the legitimate user is paying attention.
Other signals appear in the pre-authentication window, after an attacker engages with your platform but before they get inside. A breached credential being used against a login is a signal, even when the login fails. A successful defense against that attempt is not the end of the story; it is the start of the question of why this credential is now being tried, where else it might be tried, and what else the attacker may already know. New referrer sources can indicate AI-powered attacks, particularly screen scrapers.
This is why the migration toward tokenized, consumer-permissioned access (FDX APIs, user-consented data sharing) matters as a security control in its own right, whether or not the CFPB’s stalled Section 1033 rule ever compels it. Mobile app first-opens that never lead to a login attempt warrant scrutiny: Why is the app being opened by someone with no intention of authenticating? Abandoned step-up authentication challenges also deserve scrutiny, as they tend to recur.
None of these rough signals is conclusive on its own. DNS reconnaissance is noisy. The breached credential might be the legitimate user reusing a compromised password. The first open-on-mobile action might be a curious customer. The dormant account waking up might be a planned event. The signal is in the connection; when several of these patterns converge on the same identity in a short window, the picture sharpens.
A financial institution we partner with experienced multiple treasury management account takeovers (ATOs) per day, resulting in six- to seven-figure losses. The fraud team had good visibility into Phases 3 through 5: they could see authorized users being added, account information being changed, funds being withdrawn, and funds being transferred via fraudulent wire transfers.
But they had no visibility into Phase 2: how the attackers were getting in. Clients reported vishing calls in which the attackers already appeared to have access to their accounts. Something was happening upstream, and it was invisible.
The fraud team contacted cybersecurity. Together, they reviewed controls, logs, threat intelligence, and peer-shared indicators. They ruled out insider involvement, third-party compromise, breach exposure, and credential stuffing. Then a pattern emerged: a broadly targeted phishing event had triggered a spike in highly targeted impersonation vishing attacks. As they pulled the thread further, evidence of search engine malvertising and SEO poisoning surfaced.
While the cyber team took down phishing domains and traced vishing calls, the key to preventing the loss was identifying and disrupting the malvertising campaign that drove victims to attacker-controlled infrastructure in the first place.
The fraud team would never have found that on its own. This did not happen because they were incapable; rather, their tooling was looking in the wrong phase of the attack chain.
The result? Eight months without a successful ATO by that criminal group.
The earlier signals had been there the whole time. The question was whether the right teams were seeing and connecting them. The answer, it turned out, was fusion:Fraud and cyber operating as one continuous detection function rather than two siloed ones.
Here is the thesis that follows from all of this, and it requires a shift in how we think about the digital banking channel.
User education has its place. We should keep doing it. But it has a ceiling, and AI is already crafting scenarios that even sophisticated members cannot reliably distinguish from legitimate contact. The stronger play is to treat the digital banking channel itself as a continuous detection surface.
Every interaction is a signal: context, navigation, behavior, posture. The opportunity is connecting session signals to transaction decisions in real time, before the money moves. And then signals have to become action: Contextual, in-the-moment friction (e.g., “this doesn’t look like your typical transfer”) is fundamentally different from the static rule sets fraud systems have relied on for years. It is also the only real defense against the second kind of problem: When the customer is the one being manipulated, the login is clean and the step-up passes, so the tell is not who is present but whether this transfer looks like them, which is what a true detection surface can see and a static rule cannot.
When done well, you also turn potential victim touchpoints into intelligence sources by making it trivial for a user who feels suspicious to report them in-app. Then you feed that reporting directly into the same detection pipeline that watches everything else.
None of this runs on human effort alone, which brings us back to the volume problem at the top of this piece. The same agentic AI that is multiplying the attacks is now being turned on the defense. Investigation work that used to take an analyst hours (assembling the evidence, checking it against known typologies, ranking what deserves review) is being compressed into minutes by agents wired into the institution’s own systems, and several of the largest core and fraud platforms shipped exactly that capability in 2026.
McKinsey puts the productivity gain from an agentic "workforce" in financial-crime operations, one analyst supervising 20 or more agents, at as much as twentyfold, against the 15% to 20% uplift from AI that merely assists a human. Used well, it is what lets a shrinking analyst bench stay ahead of a rising signal volume. The human stops gathering evidence and starts judging the cases that matter. The detection surface produces the signals; agentic tooling is increasingly how an institution can afford to act on them.
This only works on a particular kind of platform. Specifically, it only works when the platform is architecturally unified – when the DNS signal, the breached credential check, the mobile first-open, the abandoned step-up, the dormant account waking up, and the session behavior anomaly are all visible to the same intelligence layer simultaneously.
On an assembled stack, a digital banking experience that has been stitched together from acquired components or third-party integrations, signals often do not make it past the seams. Each integration point is a place where context fails to travel.
You cannot have a continuous detection surface if your platform is architecturally discontinuous. The fusion model only scales when the underlying system is built as a single system.
The earlier signals already exist. The work is making sure they all reach the same place at the same time, so the right teams (fraud, cybersecurity, front-line support, analytics) can see and act on them together.
Everything above turns the digital channel into an apparatus that judges customers in real time. It is an assessment apparatus underpinned by a set of models owned by the risk function.
The behavioral models that decide whether a transfer looks like you, and the agents that triage what a human will never see, inherit the full discipline that the word implies: Validation before deployment, monitoring for drift, and an explicit tolerance for false positives, because friction applied to the wrong fcustomer is not free. It surfaces as abandoned transfers, support calls, and lost customers, and when it falls unevenly across a customer base, it becomes a fair-lending and UDAAP (unfair, deceptive or abusive acts or practices) exposure rather than a mere annoyance.
The supervisory ground is itself unsettled. The interagency model-risk guidance (SR 26-2, which superseded SR 11-7 this year) explicitly places generative and agentic AI outside its current scope, which means the institutions deploying these systems are, for now, defining the standard of care themselves. Treating the detection surface as a governed model estate, rather than a black box that happens to lower losses, separates a defensible program from a lucky one.
For those responsible for fraud defense or risk operations at financial institutions, I have five suggestions, in order:
Assemble your teams. Cyber, fraud, financial crime, and analytics each hold a piece of the picture. None of them sees it alone. Building the connective tissue between them, even just a minimum cross-functional review of recent loss incidents, if a more formal, funded approach for fusion is currently out of reach, is the single highest-leverage move available to most institutions today.
Map one recent loss incident across the framework. Take a recent loss and walk it across the five phases of the FS-ISAC Cyber Fraud Prevention Framework or MITRE Fight Fraud Framework. Be honest about which phases you have visibility into and which you do not. The phases where you cannot reconstruct the attacker’s path are where your missing signals live. That is your investment list.
Share intelligence with your ISAC and ISAOs (information sharing and analysis centers/organizations). Prioritize no-attribution sharing of threat paths over indicators of compromise. IOCs are perishable. Threat paths, the patterns of how attackers move through the kill chain, are far more durable, and they let your peer institutions defend themselves before they become the next case study.
Listen as much as you share. Ingest what your peers are seeing. The combination of internal fusion and external listening is what produces a durable advantage in a threat environment where attackers reuse infrastructure across hundreds of institutions before retiring it.
Govern what you deploy. The models and agents you point at this problem adhere to the same discipline as anything else that touches a customer: validation, monitoring, an explicit false-positive budget, and a check to ensure the friction lands evenly. In the current supervisory gap, you are setting that standard of care yourselves, so set it deliberately.
Fraud is as much a human problem as a technical one, and our defenses have to reflect that. The earlier signals are already there. Looking left across the full attack lifecycle, rather than reacting at the point of loss, is how we stay ahead of attackers who keep pivoting. Faster action comes from breaking silos internally, sharing intelligence externally, and treating cyber and fraud as one continuous defense function.
Loss is just latency once identity is compromised. Our work is in shrinking that latency to zero.
Sean McElroy is Chief Risk Officer of digital banking platform provider Lumin Digital. He is responsible for cybersecurity, risk management, and privacy programs to scale the next generation of cloud-based digital banking securely across global data centers. Prior to joining Lumin, he co-founded Alkami Technology in 2009 and led engineering at Computer Services, Inc. He has developed patented fintech solutions and holds multiple industry certifications including CISSP, CISM and GDSA.