Account Takeover Defence: Risk Scoring and Step-Up

Q: What is account takeover (ATO) and why is it so hard to stop in 2026?

Account takeover is when an attacker gains control of a legitimate user's account, and in 2026 it is hard to stop because it is overwhelmingly a valid-credential problem rather than a password-guessing one. Attackers no longer need to break your authentication; they arrive already holding the real username and password, sourced from infostealer malware that harvests credentials directly from infected devices and from the continuous recombination of breach corpora that leak billions of pairs. That is why MITRE ATT&CK classifies it as T1078, "Valid Accounts" — a defensive-evasion technique, because the attacker logs in exactly the way your customer does. The consequence is that a login built as a binary gate — correct password plus a one-time code — has no defence against someone who knows the password and can phish or SIM-swap the code, and real-time phishing proxies now relay OTPs the instant the user types them. Stopping ATO therefore requires moving beyond "do you know the secret" to "does this login look like the real user," which means evaluating device, network, behaviour, and velocity signals and spending additional verification only when those signals indicate risk, rather than assuming a correct password proves identity.

Q: What is risk-based (adaptive) authentication and how does it work?

Risk-based authentication, also called adaptive authentication, treats a login not as a binary gate to pass but as a risk score to compute in real time, then maps that score to an action: allow with no friction, step up to a stronger factor, or deny and alert. At the moment of login the engine ingests signals — whether the request comes from a device the account has used before (proven by a device-bound cryptographic credential, not a replayable fingerprint), the reputation of the IP and network (known proxy, VPN, Tor, or datacentre ranges), geographic plausibility including impossible-travel checks, behavioural patterns baselined against the account's own history, and velocity and graph features such as how many accounts a device is touching. Those combine into a score, and a policy layer decides the outcome. The reason this beats applying MFA to everyone is adoption, not just security: uniform friction trains users to resent and evade it, and the business eventually turns it off, whereas friction concentrated on the risky small fraction of logins is invisible to the vast majority of users and therefore survives in production. The best practice is to start with a transparent rules-plus-weights model you can explain and tune, and graduate to a learned model once you have labelled outcomes and the volume to justify it.

Q: Why is rate limiting by IP address not enough to stop credential stuffing?

Because credential stuffing at scale is run through residential-proxy pools that give the attacker a fresh, legitimate-looking IP address for every single request, which makes any defence keyed on per-IP counters trivially defeated. Credential stuffing is the automated, mass-scale trial of username and password pairs leaked from other breaches, and the attacker's economics depend on volume, so they invest in infrastructure specifically to evade naive rate limits: rotating residential IPs, human-solver farms for CAPTCHAs, and headless browsers that mimic real clients. If your only control is "block an IP after N failed attempts," the attacker simply never reuses an IP, and you also risk blocking legitimate users who share an IP behind carrier-grade NAT. The effective controls, per the OWASP Credential Stuffing Prevention Cheat Sheet, are device integrity and request-provenance signals that survive IP rotation, screening credentials against breach corpora so known-leaked passwords cannot be used at all, offering and encouraging phishing-resistant authentication, and analysing velocity and graph relationships across attempts — because stuffing lights up in the relationships between attempts (one device touching thousands of accounts) long before any single attempt looks wrong. The signal is in the pattern, not the individual request, which is the same reason graph-based methods work so well for fraud rings.

Q: What signals should feed an account-takeover risk engine?

The highest-value signals fall into four groups. First, device recognition: is this a device the account has used before, proven by a durable device-bound credential such as a cryptographic key or a bound cookie rather than a browser fingerprint an attacker can replay — a known device is the single strongest low-risk signal. Second, network and geo reputation: the reputation of the IP and ASN (known proxy, VPN, Tor, or datacentre ranges), and impossible-travel detection, where a login from one city minutes after a login from a distant one is physically implausible; geo alone is weak because of VPNs and mobile carriers but it combines powerfully with other signals. Third, behavioural signals: typing cadence, navigation path, and time-of-day patterns measured against the account's own historical baseline rather than a population average — these are powerful but privacy-sensitive, so they must be minimised, governed, and baselined per user. Fourth, velocity and graph features: how many accounts a device or IP is touching, how many devices an account is touching, and what the surrounding neighbourhood of activity looks like, which is where mass attacks reveal themselves. A strong engine also consumes external real-time events through the OpenID Shared Signals framework, so a session can be re-scored when an identity provider reports that a credential was compromised or a session hijacked, making the risk decision continuous rather than a single snapshot at login.

Q: Is multi-factor authentication enough to prevent account takeover?

No — MFA is necessary but not sufficient, and treating "we have MFA" as a complete ATO strategy is one of the most common and dangerous mistakes. Two things break naive MFA. First, factor strength: SMS one-time codes are defeated by SIM-swap attacks and by real-time phishing proxies that relay the code the moment the user enters it, and even TOTP authenticator apps, while immune to SIM swap, are relayed by the same real-time phishing kits. The only factors that genuinely resist phishing are those cryptographically bound to the origin — passkeys and hardware security keys built on FIDO2 and WebAuthn — because the credential will not release for a look-alike domain, so a phishing site cannot capture and replay it. Second, application model: MFA applied uniformly to every login trains users to approve prompts reflexively (enabling MFA-fatigue push-bombing) and pressures the business to weaken it for convenience. The effective approach is risk-based: apply step-up only when signals indicate risk, and when you do step up, step up to a phishing-resistant factor rather than a phishable one. MFA that is both phishable and uniform is largely security theatre against a modern attacker; MFA that is phishing-resistant and risk-triggered is a real control.

Q: Why is account recovery considered the weakest link in authentication?

Because account recovery is, by design, an always-available path that lets a user around every strong authentication control you built — and whatever bypasses those controls for a locked-out customer also bypasses them for an attacker. You can deploy passkeys everywhere and still lose an account through "forgot password," where the reset link goes to an email that was itself compromised, or the recovery SMS lands on a number that was SIM-swapped, or a knowledge-based question is answered from the same breach corpus that leaked the original password. Attackers know this, so a hardened login simply pushes them to always claim they "forgot" the password and attack the reset instead. The architectural fix is to treat recovery as a first-class authentication event with its own risk score rather than a customer-support convenience, and to tier its assurance: a high-value account recovering from a new device on an unfamiliar network should face more assurance than the login it is replacing, not less. That means avoiding knowledge-based questions entirely, preferring recovery through pre-registered strong factors or trusted devices, adding delays and out-of-band notifications for high-risk resets, and monitoring recovery flows for the same velocity and graph signals you apply at login. A fortress front door with a screen-door recovery path is not secure; the attacker uses the door you left open.

Q: What is the OpenID Shared Signals framework and why does it matter for ATO?

The OpenID Shared Signals framework — which includes the Continuous Access Evaluation Protocol (CAEP) and the Risk Incident Sharing and Coordination (RISC) profiles — is a standard for identity providers and relying parties to exchange security events about users and sessions in real time, such as "this credential was found compromised," "this session was hijacked," "this device was lost," or "this account was disabled." It matters for account takeover because it replaces static, snapshot-in-time trust with continuous, event-driven re-evaluation. Traditional sessions grant a token with a fixed lifetime — say thirty days — and nothing re-checks that session until it expires, which means a session that looked safe at sign-in stays trusted even after new evidence shows it is compromised, giving an attacker a long window. With shared signals, a session that was low-risk at 09:00 can be re-scored and terminated at 09:40 the moment an upstream provider emits an event indicating compromise, and your own system can publish events so downstream services react too. This turns session management from a blunt lifetime timer into a responsive control that reacts to real-world signal, and it pairs naturally with token binding to make stolen session tokens far less useful. It is the mechanism that makes "continuous authentication" concrete rather than a slogan.

Q: What is the first thing a team should do to reduce account takeover risk?

The cheapest, highest-leverage first step is credential hygiene combined with offering a phishing-resistant option, and it is the step most teams skip. Screen every password — both when it is set and at each login — against a breach corpus using a k-anonymity API so you never transmit the password itself, following the model popularised by Have I Been Pwned's Pwned Passwords; if a credential is known to be breached, it cannot stand as a sole factor, full stop. At the same time, follow NIST SP 800-63B and stop the practices that actively make things worse: forced periodic password rotation and composition rules push users toward predictable, guessable secrets, and knowledge-based security questions are shared secrets whose answers already sit in the same breach dumps, so both should be removed. Then offer passkeys so security-conscious users can move to a phishing-resistant factor immediately. This bundle removes the cheapest and most common attacks — reuse of breached passwords and basic stuffing — before you invest in a full risk engine, bot-defence tooling, or continuous evaluation, and it requires no machine learning or heavy infrastructure. Once hygiene is in place, add device and request-provenance signals to your auth endpoints, then build the transparent rules-based risk score with step-up, and only later graduate to continuous, event-driven re-evaluation and a learned model.

Satyam Kumar

Назад к блогу

security

Account Takeover Defence Architecture: Signals, Scoring, and Step-Up (2026)

By Satyam KumarSeptember 1, 202612 min read

account takeover ato defence credential stuffing risk-based authentication adaptive authentication step-up authentication passkeys fido2 webauthn phishing-resistant mfa device fingerprinting impossible travel nist 800-63b shared signals caep infostealer account recovery security fraud detection

Account Takeover Defence Architecture: Signals, Scoring, and Step-Up (2026)

Frequently Asked Questions

Поделиться статьёй

Twitter LinkedIn WhatsApp

Satyam Kumar

Founder & AI Architect, AppScale LLP

AI & Cloud архитектор. Помогаю командам строить системы, масштабируемые до миллионов.

LinkedIn GitHub

Account Takeover Defence Architecture: Signals, Scoring, and Step-Up (2026)

Frequently Asked Questions

Поделиться статьёй

Comments

Leave a comment