MFA Fatigue Investigation
Runs on alerts escalated by triage (see mfa-fatigue-triage): a user
accumulated a burst of MFA denials and a successful sign-in landed in the
same window. Triage established that the pattern is anomalous for this
user; investigation establishes whether the success was the attacker's.
The question to answer is not "did push bombing happen" — it did — but
"did it work, and what did they do with it."
Inputs
- The escalated alert (user, window, denial/success counts, source IPs, apps, user agents) and the triage report with its check outcomes.
- Raw Entra ID sign-in and audit logs. Confine all comparisons to a bounded recent window — the alert's calendar day (UTC) is usually sufficient and keeps queries cheap.
Investigation steps
Collect all four signals; the verdict is derived from them at the end.
1. Timeline reconstruction
Order every sign-in event for the user across the day. Establish: when the denial burst started and ended, where the successful login sits relative to the burst (mid-burst or terminal success is the attacker pattern; a success hours before an evening burst is not), and whether denials continued after the success (they usually stop once the attacker is in).
2. Source attribution — signal A
Compare the successful login's source against the user's own activity from the rest of the day (outside the alert window): source IP, country, ASN, and user agent.
- Signal A fires when the success came from infrastructure the user never touched that day — new country, new ASN (especially hosting/VPS ranges), or a user agent absent from their other sign-ins.
- Also record whether the denial-storm IPs and the success IP are related — same attacker infrastructure on both is a stronger story.
3. Post-authentication activity — signal B
Inspect the audit log after the successful login for actions that convert access into persistence:
- new MFA method / security-info registration
- new device registration or join
- inbox rules, OAuth consent grants, or app password creation
- privileged role activation or directory changes
Signal B fires on any of these within the day. A success with zero post-auth footprint leaves B silent.
4. Campaign check — signal C
Search the day's sign-in logs for the denial-storm and success IPs across other users. Signal C fires if the same infrastructure targeted additional accounts — this alert is then one instance of a campaign, and the report must list the other targets for separate handling.
Verdict rule
malicious— Signal A and (B or C), or B and C together: the attacker authenticated from their own infrastructure and either dug in or is running a campaign — or blended into the user's own geography (e.g. a residential proxy in the user's country and ASN) but both dug in and hit other accounts.suspicious— any other case in which at least one signal fires, or the success is terminal to the burst with no contradicting evidence: coercion is likely but unproven.benign— no signal fires and the success matches the user's own same-day infrastructure: the user eventually approved their own prompt.inconclusive— required data is missing (no audit logs, truncated window). Say precisely what was unavailable.
Output
Verdict
malicious / suspicious / inconclusive / benign, plus which of
signals A, B, C fired.
Recommended actions
Scale to the verdict — for malicious: revoke all refresh tokens and
active sessions, remove attacker-registered MFA methods and devices,
reset credentials, block the source infrastructure, and open incidents
for every account surfaced by the campaign check. For suspicious:
session revocation plus user confirmation. Never recommend auto-applying
detection changes — that is the tuning phase's job.
Evidence
The ordered timeline, the source-comparison table (success vs same-day baseline: IP / country / ASN / user agent), the post-auth audit entries, and the campaign IP-overlap results.
Reasoning
One line per signal stating fired/silent and the observation behind it, then the verdict derivation from the rule above.