Account Takeover Prevention: Eight Mistakes Banks Make and the Fix for Each
Most account takeovers don't beat your controls. They slip through the gaps between them. Here are eight gaps banks leave open, from SMS step-up to call-center resets, and the control that closes each.
SMS codes and fixed thresholds: where step-up authentication fails first
Don't ask whether you have multi-factor authentication. You almost certainly do. Ask which controls an attacker could get past if they already had the customer's password, phone number and a convincing script. Account takeover prevention for banks usually fails at the joins between controls, not because a control is missing. The first two mistakes sit in the step-up layer.
Mistake: treating SMS one-time passcodes as your main step-up. SMS is exposed to SIM swaps, number port-outs, malware that reads messages on compromised Android devices, and the simplest attack of all: a fraudster posing as the bank who asks the customer to read the code aloud. An SMS code also tells the customer nothing about what they're approving.
What to do instead: make in-app approval on a bound device your default step-up. That means a push or in-app prompt backed by a device-held cryptographic key that shows the payee and amount. Add passkeys (FIDO2) for login where your customer base supports them. Keep SMS as a fallback only, and before sending a code, check the number against a mobile operator's SIM-swap or recent-port signal where one exists in your market.
Regulators set different floors, so build to the strictest one that applies to you:
- EU and UK: strong customer authentication under PSD2 and the UK's equivalent rules requires two independent factors and, for remote payments, dynamic linking of the authentication to the amount and payee. SMS can still count as a possession factor, but supervisors and industry have been moving toward app-based methods.
- India: the RBI's authentication directions for digital payments keep two-factor authentication mandatory but allow alternatives to SMS OTP, and they expect risk-based checks.
- Singapore: MAS and industry measures have pushed major retail banks away from SMS OTP toward digital tokens. MAS's Technology Risk Management Guidelines set the broader expectations.
- Gulf states: UAE regulators have directed banks to move away from SMS and email OTPs toward in-app authentication.
- United States: there's no SCA equivalent. FFIEC guidance expects layered security sized to a documented risk assessment.
Rules change, so confirm the current text with compliance counsel before you set policy.
Mistake: stepping up on fixed rules. Challenge every login, or every transfer above a set amount, and two things happen. Customers learn to approve prompts without reading them, and attackers learn your threshold and stay just under it. Instead, trigger step-up from a risk score. Save your strongest challenge for the events that precede a loss: a new device, a new payee, a raised limit, or a change to contact details. Use number-matching on push prompts so that flooding a customer with requests until they tap "approve" stops working.
Device intelligence that never reaches the fraud engine
Mistake: running device signals in a silo. This one is common and expensive. The bank buys device intelligence, the tool correctly flags an emulator or an active screen-sharing app at login, and the result lands in an authentication log. Moments later the transaction-monitoring engine scores the payment without that context, sees a normal amount to a plausible payee, and lets it through.
What to do instead: feed device attributes into the fraud decision in real time, as model features and rule inputs alongside transaction data. When you evaluate vendors, check for these:
- Raw attributes returned through the API, not just one opaque score: device age, emulator or rooted status, remote-access tool detection, proxy or VPN use, a time zone that doesn't match the account, and recent SIM change where the operating system exposes it.
- Response times your payment flow can absorb. For instant-payment rails, aim well under 200 milliseconds.
- A persistent device ID that survives a browser reset or app reinstall without invasive fingerprinting your privacy team would reject.
- Pre-built connectors to your decisioning engine, or a clear integration path if the engine comes from another vendor.
Mistake: trusting a bound device forever. Once a device is registered, many banks never question it again. Attackers exploit this with remote-access apps: the victim installs a "support tool", and the fraud happens on the customer's own trusted phone. Fraud teams in India and Southeast Asia have reported this pattern widely, and it appears in other regions too.
Instead, treat binding as conditional. Require re-binding after a SIM change, an operating system reset or a jailbreak. Block or step up high-risk actions while a screen-sharing or accessibility-abuse app is active. Apply a cooling-off period to newly bound devices, such as lower transfer limits for the first 12 to 24 hours. Where SIM binding is a market norm, as with UPI apps in India, use it as one signal among several, not the only one.
Ignoring the session after login, and buying behavioral biometrics without a data plan
Mistake: no monitoring after login. A passed login check says nothing about what happens next. Fraud detection for online banking has to score the whole session, because the warning signs show up as a sequence of actions.
Here's a hypothetical. A session logs in from a known device. Within six minutes it changes the email address, adds a new payee, raises the daily limit and sends a transfer at the new maximum. Each step passes its own check. Together, they're a textbook takeover. The fix is session-level risk scoring that accumulates across events. A contact-detail change followed quickly by a new payee should raise the score and hold the payment, without asking the customer for yet another code.
Mistake: buying behavioral biometrics with no plan for the data. Behavioral biometrics for banks cover typing cadence, swipe pressure, navigation habits and signs of coaching such as hesitation or a call in progress. They can catch scams that every credential check misses. But deployments often stall after signature for reasons unrelated to the model.
Before you sign, settle these points:
- Legal basis and consent. Under GDPR and UK GDPR, behavioral data used to uniquely identify a person may count as special-category biometric data, which calls for a data protection impact assessment. India's DPDP Act, Brazil's LGPD and others have their own consent and notice rules. Get privacy counsel involved early.
- Data residency. Find out where the vendor processes and stores data. Payment-data localization rules in markets such as India, and sector rules elsewhere in Asia and the Gulf, may require in-country hosting.
- Retention and deletion. Agree retention periods by data type, and confirm that deletion requests carry through to the vendor's training sets.
- Feedback labels. Behavioral models drift unless you feed back confirmed fraud and confirmed-genuine outcomes. Name an owner and set a turnaround, ideally within a few days of case closure.
- Coverage. Check that the SDK covers web, iOS and Android equally, and who owns app release cycles when the SDK needs updating.
Call-center resets that undo your digital controls
Mistake: letting the contact center reset credentials on knowledge-based questions. Date of birth, address and last transaction amount are easy to find after a data breach or a little social engineering. Fraudsters know a pressured agent with a call-time target is softer than your app. They phone in, pass the questions, change the registered phone number, then take over the digital channel with full step-up rights.
What to do instead:
- Move sensitive changes out of the voice channel. For a phone number, email, device or password reset, the agent starts the request and the customer confirms it in-app on an already-bound device. If no bound device exists, route to branch or video verification with document checks.
- Score the call itself. Caller-authentication tools analyze the network path, caller ID spoofing indicators and device characteristics before the agent answers. Don't use voice biometrics as a sole factor, given how convincing synthetic voice has become. Treat it as one signal.
- Add a holding period and notify the old contact details. After a contact-detail change, block new payees and limit increases for 24 to 72 hours, and alert the previous number and email so the real customer can object.
- Give agents a scripted way to refuse. Agents need an approved way to decline or escalate without hurting their performance metrics. Otherwise urgency wins.
Self-service recovery has the same flaw. A "forgot password" path that ends in an SMS code to a number changed yesterday is the call-center gap in digital form.
Measuring alerts instead of losses, and what each fix typically costs
Mistake: judging the program by alert volume and isolated vendor scores. A team that celebrates more alerts often has more false positives, more customer friction and the same losses. Separate tools for device, behavior, call center and transactions that never share outcomes make it worse. Instead, track a few outcome metrics by channel: takeover losses as a share of digital volume, the false-positive rate on step-up, time from account change to detection, and recovery rate. Run one case-management queue, so a confirmed takeover in the contact center feeds the device and behavioral models.
Budgets vary widely by customer count, transaction volume and contract structure. The figures below are indicative typical 2026 market ranges for mid-sized institutions. Local pricing differs, so confirm with vendors.
| Capability | Mistakes it addresses | Typical 2026 pricing (USD, indicative) | Examples of vendors in the category |
|---|---|---|---|
| App-based and passkey step-up authentication | SMS-only step-up, fixed-rule challenges | $0.50–$3 per active user per year, or $100k–$800k platform fee | Transmit Security, OneSpan, Entrust, Ping Identity |
| Device intelligence | Siloed device signals, permanent device trust | $0.01–$0.10 per check; $60k–$400k per year at volume | LexisNexis ThreatMetrix, Fingerprint, SEON, Sardine |
| Behavioral biometrics | No post-login monitoring, scam coaching | $200k–$1.5M per year, usually tiered by active users | BioCatch, Feedzai, Sardine |
| Contact-center caller authentication | Weak phone resets | $0.05–$0.25 per call, or $100k–$700k per year | Pindrop, NICE, Verint |
| Real-time fraud decisioning and case management | Siloed signals, no feedback loop | $250k–$2M+ per year, plus integration services | Feedzai, NICE Actimize, Featurespace (Visa), SAS |
Integration often costs as much as licensing in the first year. Plan for roughly 30% to 100% of the first-year license for implementation, SDK work and model tuning.
A sensible order of work: fix contact-center resets and contact-change holding periods first. They cost the least and close the gap attackers find fastest. Next, connect device intelligence to your fraud engine. Buy behavioral biometrics only after you've assigned owners for consent, residency and outcome labeling. Start by listing your five most recent confirmed takeovers and checking each against the mistakes above. The gap that appears most often is your first project.
Questions buyers ask
Is SMS OTP still allowed for strong customer authentication?
In many jurisdictions, yes. Under EU and UK SCA rules, SMS can count as a possession factor, and India's RBI rules permit it alongside newer alternatives. But regulators in markets including Singapore and the UAE have pushed banks toward app-based methods, and SMS is the most attacked factor. Treat it as a fallback rather than your main step-up, and confirm current local requirements with compliance counsel.
Do we need behavioral biometrics if we already have device intelligence?
They catch different things. Device intelligence tells you whether the device and network look risky. Behavioral biometrics tells you whether the person using a trusted device behaves like the real customer or like someone being coached by a scammer. If your losses come mostly from social engineering on customers' own phones, behavioral signals add real value. If they come from credential stuffing on new devices, fix device signals and step-up first.
How long does it take to connect device signals to an existing fraud engine?
It varies widely. With pre-built connectors, plan for a few months of integration, testing and threshold tuning. Custom integrations with legacy engines can take considerably longer. Most of the time goes into feature mapping and model retraining rather than API work. Ask vendors for timelines based on your specific engine.
What is the quickest account takeover control to improve without buying new software?
Tighten the contact center and account-recovery process. Require in-app confirmation for contact-detail changes, add a 24-to-72-hour hold on new payees after such changes, and alert the old phone number and email. These are mostly policy and workflow changes, and they close a path attackers use often.
Written by the Vaultgridplus editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.