Home / Blog / Alert Triage

Intermediate Alert Triage

How to Triage an MFA Push Fatigue Alert in Entra ID

Work out whether repeated MFA prompts mean a stolen password, confirm whether anyone approved one, and contain the account using Entra sign in and audit logs.

A burst of multi factor authentication (MFA) prompts that the user did not ask for almost always means one thing: somebody else has the password. This guide shows SOC analysts in public sector organisations how to triage an MFA push fatigue alert in Microsoft Entra ID, prove whether a prompt was approved, and contain the account without breaking change control.

TechniqueT1621 Multi-Factor Authentication Request Generation
Data you needSigninLogs, AuditLogs and, with Entra ID P2, AADUserRiskEvents
Key questionWas any prompt approved, and from where?
>

Scenario: At 06:40 a policy officer in a central government department rings the service desk: her phone has buzzed with Microsoft Authenticator prompts for twenty minutes and she has just pressed "No, it's not me" and reported it. A few minutes later your SIEM raises an alert for repeated MFA failures on her account, j.okafor@example.org, from an IP address hosted outside the UK. Your job is to decide quickly whether this was a failed attempt or a successful break in.

Understand what the alert is telling you

MITRE ATT&CK tracks this behaviour as T1621, Multi-Factor Authentication Request Generation. An attacker who already holds valid credentials cannot finish signing in without the second factor, so they trigger prompt after prompt in the hope the user approves one to make the buzzing stop. ATT&CK lists APT29, a group the UK has attributed to the Russian foreign intelligence service, and LAPSUS$ among the groups that have used repeated MFA requests this way.

The important point for triage is that the password stage has already succeeded. Every MFA prompt on the user's phone means someone typed the correct password. Even if the attacker never got past MFA, you are dealing with a credential compromise.

The password is knownAn MFA challenge only appears after the first factor succeeds, so treat the credential as compromised.
Approval is the pivotOne approved prompt turns a nuisance into an incident with a live session.
Number matching helps, not curesIt stops blind taps, but a user can still be talked into reading out the number.
i

Note: Microsoft enabled number matching for all Microsoft Authenticator push notifications from 8 May 2023, and users cannot opt out. The NCSC lists push approval with number matching among its recommended MFA types for corporate services, while preferring phishing resistant methods where you can deploy them. Expect fewer accidental approvals, and watch for the follow up phone call or Teams message from someone posing as IT support asking for the number.

Find the denied and unanswered prompts

In Entra sign in logs, the result code 500121 means authentication failed during the strong authentication request, in other words the MFA step did not complete. The Status.additionalDetails field tells you why. The values that matter most for push fatigue are:

  • MFA denied; user declined the authentication: the user pressed deny.
  • MFA denied; user did not respond to mobile app notification: the prompt timed out.
  • MFA denied; user did not select the correct number: the number on the phone did not match, often because the user could not see the attacker's screen.
  • MFA denied; Phone App Reported Fraud or SuspiciousActivityReported: the user reported the prompt.
Microsoft SentinelKQL
SigninLogs
| where TimeGenerated > ago(24h)
| where ResultType == "500121"
| extend Detail = tostring(Status.additionalDetails)
| extend Country = tostring(LocationDetails.countryOrRegion)
| summarize Attempts = count(), Details = make_set(Detail, 10), IPs = make_set(IPAddress, 10), Countries = make_set(Country, 5), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by UserPrincipalName
| where Attempts >= 5
| order by Attempts desc

This gives you one row per user with the number of failed MFA steps, the reasons, and where they came from. Five is only a starting threshold. Users who lose signal on the train or change phones produce a handful of timeouts, so look at the mix of reasons as well as the count.

Read each attempt in order

For the account in the alert, expand the AuthenticationDetails column to see every step of every sign in, in time order. This shows the rhythm of the attack: prompts spaced every minute or two, a pause, then a new burst.

Microsoft SentinelKQL
SigninLogs
| where TimeGenerated > ago(24h)
| where UserPrincipalName =~ "j.okafor@example.org"
| mv-expand Step = parse_json(AuthenticationDetails)
| extend Method = tostring(Step.authenticationMethod), Result = tostring(Step.authenticationStepResultDetail), Succeeded = tobool(Step.succeeded)
| where isnotempty(Method)
| project TimeGenerated, IPAddress, AppDisplayName, Method, Result, Succeeded, CorrelationId
| order by TimeGenerated asc
+

Tip: Group the rows by CorrelationId. Several MFA failures under one correlation ID usually mean the attacker kept one sign in session open and retried, while many separate IDs mean they are starting fresh each time, which often points to automation.

Prove whether any prompt was approved

This is the question that decides everything else. Look for a successful sign in, result code 0, for the same user from any IP address that also produced the failures.

Microsoft SentinelKQL
let target = "j.okafor@example.org";
let lookback = 24h;
let deniedIPs = SigninLogs
    | where TimeGenerated > ago(lookback)
    | where UserPrincipalName =~ target and ResultType == "500121"
    | distinct IPAddress;
SigninLogs
| where TimeGenerated > ago(lookback)
| where UserPrincipalName =~ target
| where ResultType == "0"
| where IPAddress in (deniedIPs)
| project TimeGenerated, IPAddress, AppDisplayName, ClientAppUsed, AuthenticationRequirement, ConditionalAccessStatus, CorrelationId

An empty result is good news but not the end. Attackers sometimes switch IP address once a prompt is approved, so also review every successful sign in for the user since the first denied prompt, and check the device, location and application against the user's normal pattern.

Outcome 1

Benign noise

Failures from the user's own device and usual location, such as a new phone or an app stuck in a retry loop.

Confirm with the user, close with notes
Outcome 2

Attempt blocked

Prompts from an unfamiliar IP, all denied or ignored, no success afterwards. The password is still compromised.

Reset password, revoke sessions
Outcome 3

Prompt approved

A success follows the failures from the attacker's IP or another unfamiliar source. Treat as an active account takeover.

Major incident, contain now

Check the user report and the wider campaign

When the Report suspicious activity setting is enabled in the Authentication methods policy, a user who reports an unexpected prompt creates an audit event. With Entra ID P2 it also creates a high risk detection that raises the user's risk level.

Microsoft SentinelKQL
AuditLogs
| where TimeGenerated > ago(7d)
| where OperationName == "Suspicious activity reported"
| extend Reporter = tostring(InitiatedBy.user.userPrincipalName)
| project TimeGenerated, Reporter, Result, AdditionalDetails
| order by TimeGenerated desc

If you hold P2 licences, the same report appears in AADUserRiskEvents with the risk event type userReportedSuspiciousActivity, which lets your risk based Conditional Access policies act on it automatically.

Next, ask whether this user is alone. Nation state and criminal groups rarely target one mailbox when they hold a list of passwords. Pivot on the failing IP addresses across the whole tenant.

Microsoft SentinelKQL
SigninLogs
| where TimeGenerated > ago(24h)
| where ResultType == "500121"
| summarize Users = dcount(UserPrincipalName), UserList = make_set(UserPrincipalName, 20), FirstSeen = min(TimeGenerated) by IPAddress
| where Users >= 3
| order by Users desc
!

Watch out: Do not assume a user who denied every prompt was not targeted in other ways. Check the mailbox for phishing that may have harvested the password, and check whether the same IP addresses tried password spraying against other accounts before the MFA burst began.

Contain within change control

Government environments usually run strict change control, and account actions on senior officials can need sign off. Agree in advance that identity containment for a suspected compromise falls under your emergency change route, so you are not waiting for a change board while a session stays live.

Prompt approved, session active

High
Indicator

Successful sign in after the MFA burst from an unfamiliar IP, device or country.

Response

Revoke sessions, reset the password, mark the user compromised in ID Protection, and review mailbox rules, OAuth consents and new MFA methods registered since the approval.

All prompts denied

Medium
Indicator

Repeated 500121 failures with no success, often with a user report.

Response

Reset the password and revoke sessions, block the source IPs, and brief the user on why the prompts happened.

Revoking sessions invalidates refresh tokens so the attacker has to sign in again. With the Microsoft Graph PowerShell module, an authorised operator can run:

Microsoft Graph PowerShellPowerShell
Connect-MgGraph -Scopes "User.RevokeSessions.All"
Revoke-MgUserSignInSession -UserId "j.okafor@example.org"

Access tokens already issued can stay valid until they expire, so pair revocation with the password reset and, where your licensing supports it, Continuous Access Evaluation.

  1. L0
    TriageConfirm the burst, read the failure reasons and contact the user on a known number.
    SOC analyst
  2. L1
    ContainReset the password and revoke sessions under the emergency change route.
    SOC and identity team
  3. L2
    InvestigateIf a prompt was approved, review everything the session touched and widen to other targeted users.
    Incident response
  4. L3
    ReportEscalate through your department's incident process, which may include reporting significant incidents to the NCSC.
    Security lead

Close the gaps that made the attack possible

Every push fatigue alert is evidence for your Cyber Assessment Framework (CAF) work. It shows your security monitoring under objective C picked up the attack, and it points at identity controls under objective B that could be stronger.

  • Confirm number matching, application name and geographic location are shown in Authenticator push notifications.
  • Enable Report suspicious activity for all users and make sure the resulting events reach your SIEM.
  • Move privileged and high risk users towards phishing resistant methods such as FIDO2 security keys or passkeys.
  • Use Conditional Access to block or challenge sign ins from locations your department never uses.
  • Tune the 500121 analytic on the failure reasons above rather than a raw count.
  • Tell users plainly: never approve a prompt you did not start, and never read a number to someone who calls you.

Key takeaways

  1. An MFA prompt only appears after a correct password, so push fatigue always means a compromised credential.
  2. Result code 500121 plus the additionalDetails reason tells you whether prompts were denied, ignored or mismatched.
  3. The deciding question is whether any prompt was approved: look for a success from the same or any unfamiliar source.
  4. Pivot on attacker IPs across the tenant, because a single targeted user is rarely the whole story.
  5. Contain through an agreed emergency change route, then use the case as evidence for CAF identity and monitoring improvements.

Push fatigue alerts are noisy by nature, but they are also one of the clearest signs that an attacker holds a working password. Triage them quickly, answer the approval question first, and every alert becomes either a closed false positive or an early catch.

Written by Freya, Senior SOC Analyst, Government

Keep going

More in Alert Triage

BeginnerAlert Triage

How to Triage an Encoded PowerShell Alert

A step by step way to decode an encoded PowerShell command, check its parent process and related activity, and decide whether to close or escalate.

1 Nov 20268 min read