79

organizations with token replay sign-ins, per Huntress

48

rogue Entra devices registered in victim directories

19

organizations where MFA methods were wiped and replaced

90 min

roughly how long an access token outlives a reset without revocation

A password reset is not an eviction

When a Microsoft 365 account is phished, the first thing most teams do is reset the password. The campaign Huntress published on 22 Oct 2026 was built to survive exactly that. The attacker never needed a password in the first place, and within hours of getting in they gave themselves two ways back that a password change does not touch: a computer of their own registered in the victim's directory, and an authenticator app of their own registered on the victim's account.

That is the argument of this post. Against token theft, the password reset is the step that feels like remediation and is not. Eviction is a list, not a button: revoke the sessions, delete the devices the attacker joined, remove the MFA methods they added, find the inbox rules they hid, and close the sign-in flows that let them in. Miss one and the account is still theirs.

The numbers make the point. Over ten weeks, from 14 Jul to 22 Sep, Huntress tied 497 successful Microsoft 365 sign-ins at 79 organizations to one operator replaying stolen tokens from a single block of rented IP addresses. Along the way that operator registered at least 48 rogue devices in victims' Microsoft directories, and in 19 organizations wiped the victim's MFA methods and registered a new one within three hours of the first sign-in.

None of this needed a new vulnerability. Every step uses a normal Microsoft 365 feature. The attack works because the response most organizations practice was designed for the stolen password, and this one leaves the password alone.

What Huntress found in the Microsoft 365 token replay campaign

Huntress's write-up describes a campaign it calls token spraying. The name is borrowed from password spraying, where an attacker tries one likely password against many accounts. Here the attacker held a pile of stolen sign-in tokens and replayed them at volume across many organizations instead. In Huntress's words, "rather than guessing passwords, the attacker replayed sign-in tokens stolen from victims."

The replay traffic came from 193.26.115.0/24, an address block belonging to 1337 Services GmbH, a Hamburg hosting provider also known as AS210558. Huntress traced more than 200,000 failed login attempts in the campaign window to that range. Microsoft's own Entra ID Protection independently flagged four of the addresses as verified threat actor infrastructure in multiple customer tenants.

Microsoft 365 token replay campaign High
  • Reported by //

    Huntress, 22 Oct 2026
  • Window //

    14 Jul to 22 Sep 2026, about ten weeks
  • Reach //

    497 successful sign-ins at 79 organizations
  • Replay source //

    193.26.115.0/24, 1337 Services GmbH
  • Token sources //

    AiTM relays, device code phishing, social engineering
  • Persistence //

    48 rogue Entra devices, MFA methods replaced in 19 organizations
  • Goal //

    mailbox access, hidden inbox rules, payroll fraud attempts

The operator did not sit on one address. Huntress counts six main replay hosts in the /24, used one after another across four phases, often after a quiet spell of one or two weeks. Microsoft's detections named addresses one at a time, not the range. A block or a Conditional Access rule written for one address did nothing against the next. Blocking the whole /24 would have covered all six main hosts, though the operator also used a few other addresses from the same provider outside it.

Two toolsets did the replay. One, which Huntress calls Build A, was a script with the user agent node that traded stolen tokens for new ones through the Microsoft Office and Microsoft Authentication Broker client IDs, aimed at Microsoft Graph, Exchange Online, SharePoint Online and the Device Registration Service. The other, Build B, replayed browser sessions captured by the phishing relays and used a current Chrome or Edge user agent, so it looked like a person.

What the attacker wanted was email. Huntress found hidden inbox rules in compromised mailboxes with names chosen to be forgettable: a single full stop, sss, s, d. They marked mail as read and deleted it, or moved it to folders nobody opens, such as RSS Feeds and Conversation History. One rule caught any message mentioning payroll, direct deposit, bank or bonus. In that case the attacker then sent a payroll-related email to an outside accountant from the victim's mailbox, and a second rule hid the accountant's replies. Another mailbox was used to send spam to hundreds of external recipients before Entra ID Protection rated the account high risk.

None of the inbox rules came from the 1337 Services range. Each came from a different VPN or rented host, keeping the replay infrastructure and the mailbox work apart.

How token replay gets past MFA

The mechanics explain why MFA did not help here, and why the familiar response does not either.

When you sign in to Microsoft 365, you prove who you are once, with a password and an MFA prompt, and Microsoft hands your browser or app tokens that stand in for that proof. After that, the tokens are what get checked, not you. In an adversary in the middle attack, the victim signs in through a relay that sits between them and Microsoft. The relay passes the real login page and the real MFA prompt through. The victim completes both. Microsoft issues the tokens, and the relay keeps a copy. As Huntress puts it, because that token is minted after MFA, whoever holds it does not need to pass MFA again.

Device code phishing reaches the same place by a different road. The victim is persuaded to type a short code into Microsoft's legitimate sign-in page, which authorizes a session the attacker started. Huntress recorded sign-ins through the Microsoft Office device code flow from one of the replay addresses in the second phase. The victim never sees a fake page. They sign in on the real one, and hand over a valid session in doing so.

Speed is part of the design. Huntress split the campaign into two tiers: relays on one hosting provider captured sessions, and replay hosts on another used them. On 21 Jul the gap between the victim reaching the MFA prompt and the stolen session reappearing from a replay address was 12 minutes. On 22 Sep it was four. Because the two halves sit on different providers, blocking the replay range does not stop credentials being captured, and taking down a relay does not stop replay of tokens already stolen.

The 21 Jul case shows the whole sequence in under an hour:

  • 17:18 UTC: a phishing relay catches the victim at the MFA prompt.
  • 17:30: the session is hijacked from a replay address.
  • 17:43: the same session is in use from a second address at the same provider.
  • 18:08: a new authenticator app, labelled "iPhone 13", is activated on the account.

That last line is the turn. Up to 18:08 the attacker had something temporary: a stolen session that would expire or be revoked. After 18:08 they had a second factor of their own. We covered why session theft goes around MFA rather than through it in our MFA research. This campaign adds the next chapter: what the attacker does with the session once they have it.

What survives a password reset

Picture the usual response: an alert fires, someone resets the password and closes the ticket. Here is what is left standing in an account this operator touched.

Figure 01

The reset stops one thing. Four others survive it.

PASSWORD RESET

kills the stolen refresh token

  • ROGUE DEVICE

    stays joined until deleted

  • ATTACKER MFA METHOD

    passes the next prompt

  • LIVE ACCESS TOKEN

    about 90 minutes left

  • HIDDEN INBOX RULES

    keep hiding mail

A reset invalidates the stolen session. Everything the attacker built with it is still there afterwards.

The attacker's MFA method. This is the most dangerous survivor. An authenticator registered by the attacker is a legitimate second factor as far as Microsoft is concerned. Huntress is direct about it: the access "survives a password reset unless someone audits the account's MFA methods and removes the rogue one." From the second phase on, the operator was clearing the victim's existing methods through Microsoft's security info registration service and registering a new one minutes later. In 19 organizations that happened within three hours of the first sign-in from a campaign address. If the attacker later obtains the new password, by phishing again or by social engineering the help desk, the MFA step is already theirs.

The rogue device. Huntress found at least 48 computers that existed only in victims' Microsoft directories, all named DESKTOP- followed by eight uppercase hexadecimal characters. That detail is a gift to defenders: Windows names new machines DESKTOP- plus seven alphanumeric characters, so the eight character hex pattern does not match the default. On 14 Jul, in 13 of the 14 incidents Huntress responded to, a rogue device had already been registered. Huntress says a registered device "can give the attacker a more durable way to request tokens than the stolen session alone," and it stays in the directory until someone deletes it. A password reset does not delete devices.

The access token that is still alive. A reset kills the refresh token and the session cookie. Huntress saw this work: the operator's later attempts to reuse a stolen session failed with Microsoft's SsoArtifactRevoked error, code AADSTS50133. But an access token already issued keeps working until it expires, which Huntress puts at roughly 90 minutes, unless sessions are explicitly revoked. Ninety minutes in a mailbox is long enough to read the thread you are about to reply to, set a rule, and send the payment change request.

The inbox rules. A rule created during the compromise keeps running after every credential is changed, hiding the accountant's reply or the bank's fraud alert.

Revocation deserves a precise word, because it is often assumed to be part of the reset. In Microsoft Graph it is a separate action, revokeSignInSessions, which Microsoft documents as invalidating "all the refresh tokens issued to applications for a user" along with the browser session cookies. Microsoft also notes "a small delay of a few minutes" before tokens are revoked. Revocation shortens the window. It does nothing about the device or the MFA method, which are not tokens at all.

How to hunt for Microsoft 365 token replay in Entra logs

Start by assuming the response playbook is the first finding. Then look in the places this operator left marks.

Sign-in logs. Build A traffic has a profile: OAuth2 token requests through the Microsoft Office or Microsoft Authentication Broker client IDs, with a user agent of node or a bare Mozilla/5.0, from hosting infrastructure. Huntress recommends hunting for exactly that, and for sign-ins to the Device Registration Service from hosting addresses. Expect some legitimate automation and focus on sign-ins that touch mailboxes or device registration.

Conditional Access blocks. Huntress makes a point that is easy to miss. A Conditional Access block, error AADSTS53003, happens after authentication has succeeded. A block means the attacker held a valid credential or token and a policy stopped them at the last step. Treat it as a compromised account that happened to be caught, not as a save.

Rank by validity, not volume. In the third phase, one replay address logged 1,052 blocked sign-ins, all from four organizations where Conditional Access was doing its job. Its successful sign-ins landed in more than two dozen other organizations, with no overlap. A responder sorting tenants by failed attempts would have looked at the four that were fine.

Devices and MFA methods. List Entra devices with names matching DESKTOP-, Device- or Work Pc followed by eight hexadecimal characters, especially any registered from a hosting address. Huntress warns that the pattern is rare but not unique to this campaign, so confirm before deleting. In the audit log, look for an account whose MFA methods were cleared and a new one registered minutes later.

Inbox rules. Look for rules with one to three character or punctuation-only names, rules that move mail to RSS Feeds, RSS Subscriptions or Conversation History, and rules keyed on payroll or banking terms. Huntress found four rules created through the Outlook REST API rather than the Outlook client, two within minutes of the first sign-in.

Figure 02

Scripted token replay through Office clients, and device registration from the campaign range

KQL

// Microsoft Office and Microsoft Authentication Broker client IDs, per Huntress
let replayClients = dynamic(["d3590ed6-52b3-4102-aeff-aad2292ab01c", "29d9ed98-a469-4536-ade2-f981bc1d605e"]);
// Device Registration Service resource ID, per Huntress
let deviceRegistration = "01cb2876-7ebd-4aa4-9cc9-d28bd4d359a9";
union isfuzzy=true SigninLogs, AADNonInteractiveUserSignInLogs
| where TimeGenerated > ago(30d)
| where AppId in (replayClients)
| where UserAgent in ("node", "Mozilla/5.0") or ResourceIdentity == deviceRegistration
| extend CampaignRange = ipv4_is_in_range(IPAddress, "193.26.115.0/24")
| summarize SignIns = count(), Results = make_set(ResultType), Resources = make_set(ResourceDisplayName), Addresses = make_set(IPAddress), Agents = make_set(UserAgent) by UserPrincipalName, CampaignRange
| order by CampaignRange desc, SignIns desc
A starting point to tune against your own telemetry, not a production rule. Any row with CampaignRange true is an incident. Every other row needs a reason, especially result code 0 or 53003, and any sign-in to device registration from a hosting provider.

If you confirm a hit, revoke sessions and reset the password together, then clear the device, the method, the rules and any new enterprise application consent, which Huntress lists as another way to stay. Then find out what the mailbox sent while it was not the user's. The payroll email is the part that costs money.

How we test Microsoft 365 session security

On a sanctioned engagement, we treat Microsoft 365 the way this operator did: as a set of sessions and enrollment flows, not a login page. The question we answer for the client is not "can we phish a password" but "if a session is stolen, what can the holder do with it, and does anyone notice?"

At a high level, we start with which sign-in flows the tenant still allows, including device code flow and legacy protocols, and whether Conditional Access covers every user, app and client type or leaves gaps for service accounts and old clients. Those gaps are where we spend our time.

Then, with a test account the client provides, we work out what a stolen session could build. Can that account register a device without a compliant or managed device requirement? Can it add a new MFA method without a fresh strong authentication, or after its existing methods are cleared? Can it create an inbox rule or consent to an application? Each answer is a persistence path, and each path is something the client's playbook has to undo.

Last, we test the response. We give the client's team the indicators of a stolen session and watch what the playbook removes. If the account comes back clean after a password reset but our registered device and our authenticator are still there, that is the finding, and it matters more than the phish that started it. Our password spraying and MFA fatigue guide covers the Conditional Access and detection checks that sit next to this work.

Two small tools help. Our free JWT decoder shows when a token from your own test tenant expires, which makes the access token window concrete for a response team. The phishing analyzer on our open source page pulls the links out of a suspected lure so the relay domain can be blocked and hunted.

What is still coming for token theft

Expect more of this, for three reasons.

The first is that token theft is cheap to scale. This campaign shows one operator with a script and a rented address block working through stolen sessions from dozens of organizations at once. The defensive answer to password spraying took years to become routine. The answer to token spraying is still being written into most playbooks.

The second is that persistence is now part of the attack, not an afterthought. Registering a device and an authenticator took the operator minutes. Those are self-service features that exist so employees can enroll a new phone without calling the help desk. Every self-service enrollment flow is also an attacker enrollment flow, unless it demands a strong authentication that a stolen session cannot provide.

The third is the gap between what Microsoft offers and what tenants switch on. Microsoft recommends organizations "get as close as possible to a unilateral block on device code flow," and documents a Conditional Access policy that does it. Token Protection is a Conditional Access control that only accepts sign-in session tokens bound to the device they were issued to, so a copied token cannot be used from another machine. It is generally available for native applications on Windows and covers Exchange Online, SharePoint Online and Teams. Coverage for browsers is still limited, which matters because the relays in this campaign stole browser sessions. Binding tokens to devices narrows the window this operator used. It does not yet close it, and it only helps in tenants that turn it on.

What to do today

  1. Rewrite the compromised account playbook. Revoke sessions and reset the password together, then remove unknown Entra devices, unknown MFA methods, hidden inbox rules and new application consents. A reset on its own leaves all four.

  2. Block device code flow unless you need it. Use the Conditional Access authentication flows condition, start in report-only mode, and keep any exceptions to named, restricted accounts.

  3. Close the Conditional Access gaps. Require MFA or block for all users, all cloud apps and all client app types, and treat an AADSTS53003 block as a valid credential that needs a response.

  4. Hunt this week. Run the KQL above, list devices named DESKTOP- plus eight hex characters, and look for MFA methods cleared and re-registered within minutes.

  5. Make MFA enrollment and device registration ask for more. A stolen session should not be enough to add an authenticator or join a device.

  6. Pilot Token Protection for native apps. Start with Exchange Online and SharePoint Online on managed Windows devices, in report-only mode first, as Microsoft advises.