The distinguishing evidence is the sequence across both tenants, not just an anomalous login. In organization one, correlate the super-admin sign-in with MFA denials followed by approval, residential-proxy IP/ASN, unfamiliar device identifiers, device-registration or join events, authentication-method changes, Primary Refresh Token issuance, and privileged directory activity. Then look for mailbox, SharePoint, OneDrive, or Graph searches and bulk access focused on employees, contacts, and relationship history. Link that to subsequent messages or lures aimed at those exact external contacts. That chain—privileged takeover, relationship-data collection, then tailored external targeting—is the trust pivot. One strange login is ordinary compromise; this is identity compromise being converted into a stolen introduction.
Containment must remove the attacker’s new keys, not merely close existing sessions. Disable the affected privileged identities; delete unauthorized registered devices and authentication methods; invalidate refresh tokens and device-bound credentials; reset passwords; require clean enrollment with phishing-resistant MFA; and review every privileged-role assignment, Conditional Access change, OAuth consent, application credential, federation setting, Temporary Access Pass, inbox rule, forwarding rule, and delegated mailbox permission touched during the exposure window. Rotate secrets for applications and service accounts the administrator could access. Quarantine the impersonated mailbox, preserve Entra and Microsoft 365 audit evidence, block the phishing infrastructure across both organizations, and warn identified relationship contacts through a separate trusted channel.
For organization two, a personalized lure containing stolen relationship context proves targeting, not tenant compromise. Even credential submission only proves disclosure unless there is evidence the credentials or tokens were accepted. Treat the organization as compromised when telemetry shows successful attacker authentication, token issuance or replay, access from an attacker-controlled device or proxy, Graph/Exchange/SharePoint resource access, unauthorized device or MFA registration, OAuth consent, mailbox-rule creation, directory enumeration, or data download. Correlate Entra sign-in IDs, device IDs, token/session timing, IP infrastructure, user agents, and subsequent resource audit records. A failed phishing login followed by no token or resource activity remains an attempted intrusion; an accepted token followed by one successful API call closes the initial-access chain.
The attribution label does not change that threshold. “Star Blizzard” may guide infrastructure hunting, but the incident decision should rest on accepted authentication and resource access—not branding, lure quality, or the number of MFA prompts.