Analysis

Risk-Based Authentication and What to Do When a Session Risk Score Rises 

Clara Santos Product Marketing Manager
20 November, 2026 12 min

Frequently asked questions Risk-based authentication decides, at each point in a session, whether an additional proof of identity is needed based on the risk that session is showing, instead of always requesting it at the same fixed checkpoints. It starts from something every fraud team has learned first-hand: a successful login says very little about what will happen in the minutes that follow. The credentials can be correct and the face can belong to the account holder, and the transfer ordered moments later can still be in someone else’s hands, or in the hands of the same person acting under deception. 

The most recent data points in that direction. In Mexico, claims for possible fraud at commercial banks rose 18% between January and May 2026 compared with the same period the year before, according to the Comisión Nacional para la Protección y Defensa de los Usuarios de Servicios Financieros (CONDUSEF), the public body that handles complaints from users of the Mexican financial system. In Europe, the latest joint report by the European Banking Authority and the European Central Bank describes a deeper shift. Strong customer authentication continues to stop the fraud it was designed for, while the fastest-growing schemes either target transactions exempt from it or lead users themselves to authenticate the fraudulent transaction. 

Many institutions already calculate a risk score for every session. What usually remains less settled is what to do with that number, when it is worth reading, and who has the final say when it rises. 

What is risk-based authentication? 

Risk-based authentication is an authentication model that decides whether to request an additional proof of identity based on the risk at each point in the session, instead of always requesting it at the same checkpoints. It is also known as adaptive authentication, the usual term in access control. Here, however, risk is assessed throughout the entire operation, not only at login. 

The traditional model applies multi-factor authentication at fixed checkpoints. Users are verified when they log in, verified again before a large transfer, and in between it is assumed that nothing has changed. Adaptive authentication keeps the same verification mechanisms, such as facial biometrics or a second factor, and changes when they are applied. At each point in the session, the question is whether new proof is needed, and the answer depends on what the session is showing. 

The European framework already thinks along these lines. The PSD2 regulatory technical standards on strong customer authentication, set out in Commission Delegated Regulation (EU) 2018/389, allow a remote transaction to be exempted from it when a real-time risk analysis classifies it as low risk, within certain thresholds. They also require that analysis to combine several factors into a risk score for each transaction. The idea of scoring risk has been in regulation since 2018, and since then what an institution can observe within a session to feed that score has grown considerably. 

What is a session risk score? 

A session risk score is a numerical indicator that summarises the risk of an interaction at a given moment, calculated from several combined signals. The higher the value, the higher the risk. 

The signals behind it come partly from device intelligence: the state of the device (whether it is rooted, whether it is an emulator, whether there is active malware or remote control tools), the network it connects from, and the location and timing of the session compared with that account’s usual history. When an institution adds behavioural biometrics, it can also observe how the person types and navigates and compare it with their usual behaviour profile. 

Two nuances change how the score should be read. It is recalculated throughout the session, and legitimate customers notice nothing as long as their behaviour matches what is expected. And no single signal condemns a session. Plenty of people use a VPN for perfectly legitimate reasons, and an unusual time of day may simply mean someone is travelling; what carries information is how the signals combine. 

That is why I insist on separating the score from the decision. A 75 describes a session but does not say what to do with it, and two institutions looking at the same 75 may want very different responses depending on their product, their risk tolerance and their regulatory environment. 

When to assess session risk: four key moments 

The score is updated continuously, but there are moments when reading it carries much more weight, because they tend to come just before money leaves the account. In my view, any response policy should cover these four. 

Adding a new payee 

In an account takeover, the attacker needs somewhere to send the funds, and adding a payee is usually the step that comes first. European regulation already requires strong customer authentication when a list of trusted beneficiaries is created or amended, but that only proves the factors were presented, not that the person presenting them is the account holder. Reading the session adds whether the context in which they are used makes sense, for instance whether the device shows signs of remote control or whether the way the person interacts with the app has changed compared with a few minutes earlier. 

Changing a phone number or email address 

Changing contact details redirects second factors and notifications, so the account holder stops receiving the alerts that might warn them. Some types of banking malware can also intercept verification text messages without the user noticing, which means the change may arrive validated with a legitimate code. It is a quiet moment with no money moving, and that is exactly why it should be treated as a sensitive operation. 

First transfer to an unknown destination 

This is where bank transfer fraud concentrates, and it is also the hardest moment to interpret. The EBA and ECB report, based on 2024 data, shows that credit transfer fraud losses in the European Economic Area reached €2.2 billion that year and that users bore around 85% of them, mainly due to scams that led them to initiate the transaction themselves. 

There are two scenarios. When someone operates the victim’s device remotely, the session usually gives it away through active remote access tools, navigation that does not match the account holder’s or unusual timing on the confirmation screen. When it is the customer who operates, persuaded by a fraudster, the session reveals less, because the credentials, the device and the behaviour are all genuinely theirs. In those cases, a complete response needs information the session does not contain and the institution already holds in its own systems: who the customer has paid before, which amounts are normal for them and whether this destination is appearing for the first time. 

Raising limits 

Raising a transfer or withdrawal limit is a classic sign of preparing to empty an account. On its own, it may be a perfectly legitimate customer decision. Read alongside the previous steps (a new device, a changed email address, a newly added payee), it forms a sequence that weighs far more than any single step. My recommendation is that the response policy considers that full sequence, not just the isolated event. 

From risk score to decision: allow, quarantine, challenge or block 

With the score calculated and the moment identified, what remains is deciding the response. The approach I advocate is proportional: it reserves visible friction for the cases where it pays off, with four possible decisions ordered from lowest to highest risk. 

Risk level Decision What the user sees 
Low Allow Nothing. The transaction proceeds as normal. 
Medium-low Quarantine Nothing. They keep operating, and the fraud team reviews the session afterwards. 
Medium Challenge An additional verification, usually facial biometrics, before continuing. 
High Block The session is stopped or the threat is neutralised immediately. 

Only one of the four decisions, the challenge, visibly interrupts the customer experience, and that explains much of the balance between security and friction. In addition, when several alerts with different levels are triggered within the same session, the most restrictive decision prevails: if one rule says allow and another says block, the session is blocked. The reason is the asymmetry of costs, because in security, letting through something that should have been stopped costs far more than acting when it perhaps was not necessary. 

When to quarantine a session 

Quarantine fits when there is enough evidence to investigate, but not enough to interrupt. The customer keeps operating, the session is flagged internally and the fraud team reviews it afterwards. It is the decision most easily forgotten when designing a policy, because it is invisible; yet it allows institutions to learn from the doubtful case without making the legitimate customer pay for it. 

When to trigger a dynamic challenge (step-up authentication) 

At medium risk, a dynamic challenge is triggered. The dynamic challenge is the mechanism that carries out a new act of verification (for example, a facial re-verification or a one-time code) that is only requested when the context calls for it. 

A dynamic challenge can appear at any point in the session: at login, before confirming a transfer or in the middle of a change of details. When in doubt, the institution re-verifies instead of ending the relationship, and the legitimate customer passes the verification and carries on. However, the challenge confirms who is on the other side, not whether that person is being deceived. 

When to block 

At high risk, the response is immediate: the session is stopped according to the institution’s rules. The difference from a system that only raises alerts lies in the timing, because the transaction can be stopped before it is completed, while the money has not yet moved. 

How to set thresholds without penalising legitimate customers 

Thresholds are configured as rules in the risk engine, and each rule has three components: a business objective, a set of conditions and an action. Here is an illustrative example, with figures that should not be read as a recommendation. 

  • Objective. Protect transfers. 
  • Conditions. Session risk score above 80 and banking malware detected on the device. 
  • Action. Block. 

Another institution might decide, faced with the same scenario, to trigger a dynamic challenge instead of blocking. Neither is wrong, because each responds to different products, customers and risk appetites. 

From this framework I draw three practical criteria. The first is to start from the business objective rather than the number, because a threshold without an objective ends up being adjusted by guesswork. The second is to combine the score with a specific signal whenever possible, since a rule that requires both a high score and active malware is far more specific than one that looks only at the score. The third is to measure friction as well as fraud, because how many challenges a legitimate customer receives is as relevant an indicator as how many fraudulent transactions are stopped. 

History also matters for accuracy. Someone logging in at dawn from a new city may be a traveller or an attacker, and what makes the difference is knowing that account’s pattern. With a new customer, device, network and location signals provide information from the very first session, because they do not depend on the customer’s history (an emulator or a remote control tool is suspicious on any account), while the usual behaviour profile is built up through use. Feeding the outcome of each investigation back into the system, whether confirmed fraud or false positive, helps refine the models over time. 

What the rules engine decides and what the analyst decides 

The risk engine issues a recommendation in real time by applying the logic defined by the institution, and it can end the session when the institution’s rules provide for it. That recommendation reaches the institution’s systems, which decide how to carry it out within their own processes, whether by ending the session, triggering a challenge, escalating the case to manual review or applying any other measure set out in their policies. Whoever holds the relationship with the customer and answers to the supervisor retains the authority to act on their account, and the technology gives them the risk reading to do so with sound judgement. 

The analyst comes in afterwards, and their job is to understand what happened. To do that, they need to reconstruct the session: which screens the user visited, in what order, how long they spent on each, when the alerts were triggered and what state the device was in. Consider a session in which the user spends barely two seconds on the transfer screen when they usually take much longer to review it, while the device shows signs of remote control. With a timeline of the session in front of them, the analyst can see the full sequence and explain why the recommendation was issued. 

That traceability goes beyond investigation, because it helps reconstruct the sequence of the session for a customer who files a complaint, for internal audit or for the supervisor. For compliance teams, being able to reconstruct the reasoning behind a decision often carries as much weight as the accuracy of the model itself. 

Frequently asked questions 

It is an authentication model that decides whether to request an additional proof of identity based on the risk at each point in the session, instead of always requesting it at the same checkpoints. It keeps the same verification mechanisms and changes when they are applied.

It is an indicator from 0 to 100 that summarises the risk of an interaction at a given moment, calculated from several combined signals such as device state, network, location, session timing and, if enabled, behaviour. The score describes the risk, and the decision depends on each institution’s rules.

In quarantine, the customer keeps operating as normal and the fraud team reviews the session afterwards. Blocking stops the session or neutralises the threat immediately. Quarantine is reserved for medium-low risk and blocking for high risk.

Device, network and location signals provide information from the very first session, because they do not depend on the customer’s history. The usual behaviour profile is built up through use and becomes more accurate as sessions accumulate. 

The platform issues a recommendation based on the rules each institution defines. Final execution rests with the institution and its analysts, who decide within their own processes whether to end the session, trigger a challenge or escalate the case to manual review.

Ready to protect your users?

Discover how Facephi's biometric technology can safeguard your identity verification process.

Facephi Facephi Identity Platform Onboarding Authentication UX Consultancy Facephi Builder Facephi Central Services Fraud Intelligence Platform Identity Fabric KYB Platform Teseo Identity Wallet IDV Suite Cuentas Mula Behavioural Biometrics Linkedin YouTube X Facebook
Secret Link