Token theft and AiTM phishing: MFA isn't the finish line
Adversary-in-the-middle phishing kits don't steal your password, they steal the session you earn by passing MFA. Here's how the attack works, why it keeps coming back, and the controls that make a stolen token far less useful.
- What: AiTM kits proxy the real Microsoft sign-in page. The victim completes MFA for the attacker, and the kit keeps the session.
- Why it matters: a replayed session gets past MFA, so "we have MFA" isn't the finish line.
- Best fix: phishing-resistant sign-in (FIDO2 keys, Windows Hello for Business, certificates), starting with admins.
- Also do: block device code flow, require managed devices for sensitive apps, and use token protection where it's supported.
- Be honest about gaps: token protection and continuous access evaluation both have limited coverage today. Plan to detect and respond, not just prevent.
What AiTM phishing actually is
In a classic phishing attack the criminal collects your password, and MFA is what stops them using it. AiTM (adversary-in-the-middle) phishing changes the job. The kit runs a reverse proxy between the victim and the genuine Microsoft sign-in page. The victim sees the real thing, types their password, approves the MFA prompt and carries on, while the proxy captures the username, password, MFA code and, crucially, the session cookie issued once sign-in succeeds, as Trend Micro describes it.
An attacker who replays that cookie takes over the account even though MFA was switched on. That's the point to hold on to: the attacker never has to beat MFA, they borrow the result of someone else passing it.
Why it keeps coming back: Tycoon 2FA
Tycoon 2FA was a subscription phishing-as-a-service platform, active since August 2023, built to bypass MFA this way. Investigators say that at its peak it accounted for about 62% of the phishing attempts Microsoft blocked. In March 2026 Microsoft led a disruption alongside Europol-coordinated law enforcement in Latvia, Lithuania, Portugal, Poland, Spain and the UK, seizing 330 domains. Europol said the platform produced tens of millions of phishing emails a month and had given criminals access to nearly 100,000 organisations.
Did the takedown solve it? The evidence is mixed. SecurityWeek, as relayed by Pulse, reported that activity dropped by roughly 75% and then returned to its earlier level within days, although that report gives little data to back the rebound. Trend Micro, writing in the same week, was more cautious: operators are known to adapt and move to new infrastructure, and sessions and credentials stolen earlier remain in circulation. The practical reading is that one takedown doesn't retire the technique, and it's worth reviewing sign-in logs from before it, not just after.
A newer twist: device code lures
Huntress's September 2026 write-up of a kit called Knight Office shows how these kits keep evolving. The lure is a DocuSign-style email with a spoofed sender. The link bounces through a tracking URL to a compromised website, which shows a fake SharePoint or Teams notice and a device code to copy. The victim is then sent to Microsoft's real device sign-in page, enters the code and approves the MFA prompt, and the kit captures the session tokens.
What happens next is the part defenders miss. According to Huntress, the operator replays the tokens from data centre infrastructure, enrols a rogue device in Entra ID and binds a Windows Hello for Business key to the account. That gives them a way back in even after the stolen tokens are revoked. Everything the victim sees is genuine Microsoft, including a warning about entering codes from untrusted sources, which is why user awareness alone is a weak defence here.
What actually helps
None of these is a silver bullet. In rough order of value:
- Move to phishing-resistant sign-in. This is the only control here that removes the problem at source. Entra Conditional Access has a built-in "Phishing-resistant MFA" authentication strength that accepts Windows Hello for Business or a platform credential, FIDO2 security keys and certificate-based authentication. Microsoft Authenticator phone sign-in sits in the passwordless strength, not this one. The UK NCSC's April 2026 passkey guidance makes the same direction of travel: passwords plus one-time codes remain phishable by design, and FIDO2-based credentials are judged at least as secure as traditional MFA against common attacks. Start with admins and other high-impact accounts.
Three caveats. Conditional Access evaluates the strength after the first factor, so users must have registered the methods first. It needs Entra ID P1, and it can't sit in the same policy as the "Require multifactor authentication" grant. And the NCSC paper is aimed mainly at consumer-facing services and says it isn't meant for enterprise risk assessment, so treat it as supporting evidence rather than a design document. The NCSC also stresses securing account recovery and fallback, because weak resets reintroduce the risk. - Block device code flow. Microsoft recommends getting as close as possible to a blanket block and allowing it only for well-documented, secured cases such as legacy tooling that can't be updated. That removes the route Knight Office-style lures depend on. Build the policy in report-only mode first, exclude your emergency-access accounts, and audit any other exclusions regularly.
- Require a managed device for sensitive apps. Intune reports device compliance to Entra ID and Conditional Access can grant or block on it. Note that Microsoft's own sample policy accepts MFA, a compliant device or a hybrid joined device as alternatives, so MFA alone satisfies it. If you want device status to count against a token replayed from an attacker's machine, it has to be a requirement in its own right for the apps in question. That's my reading rather than a Microsoft recommendation for AiTM. Expect certificate prompts for some browsers on macOS and mobile, keep break-glass accounts excluded, and run report-only first, as a misconfigured policy can lock admins out.
- Turn on token protection where it's supported, and know its limits. It binds sign-in session tokens to the device, so a stolen token can't be replayed from another one. Native apps for Exchange Online, SharePoint Online and Teams on Windows are generally available. Browser support is still preview and limited to selected web apps reaching Azure Resource Manager. Microsoft itself describes it as reducing replay rather than a complete fix. Because AiTM kits capture sessions through the browser, where coverage is currently thin, treat it as one layer. Pilot it first, review both interactive and non-interactive sign-in logs, and check the platform deployment guides for the current status on Apple devices, as Microsoft's own page is inconsistent on that point.
- Shrink the window, and don't over-trust CAE. Continuous access evaluation can cut off a token before it expires, but only where the resource and the client both support it. ERNW's research, reported by Insinuator, found only 33 of 740 checkable first-party resources returned the CAE claim, that clients choose whether to ask for CAE tokens, and that an access token already issued can keep working for up to about 90 minutes after you disable the account or revoke refresh tokens. They also found that setting a sign-in frequency control brings CAE tokens back to the normal 60 to 90 minute lifetime. Their revocation timings come from only three measurements per resource and event, so treat those as directional.
Spotting it in your sign-in logs
Huntress lists signals that point to replay rather than a normal sign-in:
- Authentication after MFA from an IP address that differs from the one the user signed in from.
- Sign-ins from devices you don't recognise.
- Successful access with no failed password attempts, which suggests a replayed token rather than guessing.
- The
python-requestsuser agent on token replays, and theDsreguser agent during Windows Hello for Business enrolment. - Newly registered Entra devices or new Windows Hello for Business credentials.
That last one is worth an alert of its own. Persistence is the step that turns a one-off phish into a lasting foothold.
If you think a token has been stolen
- Revoke sessions and reset credentials, but don't stop there. In the Knight Office case the operator had already enrolled a device and added a Windows Hello for Business key, so revocation alone didn't lock them out.
- Review and remove what was added. Huntress's guidance is to check newly registered Entra devices and Windows Hello for Business credentials, remove unauthorised authentication methods, and disable synchronised accounts in on-premises identity systems.
- Allow for a tail. Access tokens already issued can keep working for up to about 90 minutes, so watch the account's activity through that window.
- Look at what the session did. Check mail access, new inbox rules and file downloads. If a mailbox was in play, the payment-fraud controls in the BEC post apply.
The bottom line
MFA is still essential, but it's no longer the end of the conversation. The direction of travel from both Microsoft and the NCSC is phishing-resistant sign-in plus device-bound tokens, and neither is complete today. So combine the layers, tighten the lifetime of what an attacker can steal, and watch for the signs that a session has been replayed.
Sources
- Trend Micro: Tycoon 2FA takedown
- Help Net Security: Authorities pull plug on Tycoon 2FA phishing-as-a-service platform
- Pulse (reporting SecurityWeek): Tycoon 2FA fully operational despite takedown
- Huntress: Inside Knight Office, a new M365 AiTM phishing kit
- Microsoft Learn: Conditional Access authentication strengths
- Microsoft Learn: Block authentication flows with Conditional Access
- Microsoft Learn: Require a compliant device with Conditional Access
- Microsoft Learn: Token protection in Conditional Access
- CSO Online: Offer customers passkeys by default, UK's NCSC tells enterprises
- Insinuator (ERNW): Token theft in Microsoft Entra ID, part 2, continuous access evaluation
- Microsoft 365 for IT Pros: Entra ID continuous access evaluation and Microsoft 365