A reverse proxy that finishes the employee’s real MFA approval is not a clever bypass of cryptography. It is a completed login under attacker control. According to Unit 42 at Palo Alto Networks on November 20, 2026, Com-affiliated actors including Bling Libra and clusters CL-CRI-1264 (Vortex) and CL-CRI-1274 (Cinder) posed as internal IT support, steered staff to organization-tailored adversary-in-the-middle sites, relayed workforce passwords and MFA approvals in real time, then used the resulting Microsoft 365 sessions for mailbox, OneDrive, and SharePoint exfiltration plus internal extortion. Device-bound, origin-bound authentication with no transferable secret would have stopped that proxied workforce login before any session existed for the caller.

What Unit 42 logged in the Com AiTM vishing wave

Unit 42 described sustained Com-affiliated vishing from at least August 2026, with earlier related data-theft and extortion work earlier that year and a clear acceleration into the autumn. Callers claimed to be internal IT or helpdesk support. The script pushed employees toward lookalike portals built for that organization, not generic spray pages. Apex domains and short-lived subdomains impersonated SSO portals, IT support pages, and even passkey or MFA registration flows so the destination felt routine on a corporate workstation or a personal phone.

An adversary-in-the-middle attack is a reverse proxy between the user and the real identity service so the password and the second factor both complete under attacker control. Unit 42 stated those servers “act as adversary-in-the-middle (AiTM) proxies that pass user credentials and multi-factor authentication (MFA) approvals to the attackers in real time.” The employee typed a password. The employee approved MFA. The proxy forwarded both and kept the authenticated Microsoft 365 session.

Targeting shifted during the year. By September 2026, CL-CRI-1264 and CL-CRI-1274 were identified as pressure moved from financial services toward healthcare. October kept healthcare and financial services near the top while Australian targeting rose. Public reporting does not name individual victim companies, publish exact victim counts, per-victim volumes, or ransom payment amounts, so the picture stays at industry and geography rather than a clean victim list.

The same Com-affiliated ecosystem also showed a parallel harvest path under CL-CRI-1192 (Helix): a sideloaded browser extension that pulled authentication cookies and MFA tokens, including via QR scans. That path sat beside the core vishing-plus-proxy clusters rather than replacing them. The enabling step on the main clusters remained a coached workforce sign-in that a stranger could finish from a phone call and a tailored proxy. The full attack narrative lives on the companion write-up at legacymfa.sucks.

Why transferable MFA still finished the proxy ceremony

The initial access stage was credential theft through live AiTM relay. Operators did not need to break the second factor offline. They needed the employee to complete a password and an MFA approval on a page the proxy controlled. That is the same industrial class as coaching someone through a fake login: a transferable factor moves through an attacker-controlled origin while the worker still believes IT is helping.

Centralized cloud collaboration stacks made one successful ceremony expensive. A single workforce identity often reached mail, files, and chat after one accepted sign-in. Unit 42 noted victims commonly used centralized SSO and cloud collaboration platforms, which is why the operators invested in organization-tailored pages instead of one-size-fits-all lures.

Persistence split into two documented patterns after that first accepted login. Attackers behind CL-CRI-1264 enrolled an unauthorized MFA device within minutes of authenticating, including reuse of a device named Pixel 10a across accounts. That move turned a one-time proxied login into a factor the operators could satisfy later without another perfect vishing call. CL-CRI-1274 took a different cut. Unit 42 reported those operators used vishing for initial access and avoided re-prompting for Duo MFA by replaying a durable signed-in session token obtained through the proxy infrastructure. The MFA prompt never returned because the durable session already said the user was signed in. Public detail does not claim identical MFA product behavior for every cluster, so the Duo keep-me-signed-in case stays tied to the cluster Unit 42 described.

Data access sat entirely after authentication. With Microsoft 365 sessions in hand, operators used Python tooling built on httpx and requests plus the Microsoft Graph SDK to pull mailbox, OneDrive, and SharePoint content at script speed. Extortion notes went out over Microsoft Teams or email toward Tor-hosted negotiation sites, so pressure landed inside the same collaboration tools the business already trusted. Occasional large CRM or recruiting record theft appears in the reporting without published totals.

Stage What operators held Login MFA still in play?
AiTM proxy login Password + live MFA approval Yes, transferable ceremony
Rogue MFA enrollment Stolen session, then new factor No, session already issued
Keep-me-signed-in replay Durable session cookie No, post-sign-in token abuse
Graph exfiltration Live M365 session No, cloud access via token

Password-plus-approval stacks fail here because both halves are relayable. The second factor proves someone can approve a prompt or type a code. It does not prove the private key never left a hardware boundary on the legitimate origin.

Where device-bound phishing-proof MFA stops this campaign

Phishing-proof device-bound MFA changes the initial-access math. There is no password to type into a lookalike portal and no OTP, SMS code, email code, or push approval a reverse proxy can forward. The private key stays on the enrolled device. The signature is origin-bound, so a spoofed host is the wrong party and cannot complete a usable workforce login. That is prevention, not detection: the AiTM kit never receives an authenticated Microsoft 365 session from that path.

The same property closes the cheap remote playbook Unit 42 documented. Callers can still sound like IT. They can still send a tailored URL. They cannot harvest a transferable second factor or a password the proxy can replay to the real identity provider. Corporate workstations and personal phones both lose the relay surface when login only accepts hardware-bound signatures for that origin.

Stages after a session already exists are different work. MFA 2.0 does not retract a cookie or OAuth session that malware or a completed legacy login already placed in attacker hands. Graph scripts against mailbox and SharePoint, keep-me-signed-in replay, and internal Teams extortion are containment problems once the token is live: revoke sessions, disable attacker-registered authenticators, hunt Graph abuse, and shorten what a single stolen session can still reach. Those controls matter. They are not substitutes for removing the phishable login that produced the session in this campaign.

Unit 42 observed unauthorized devices added within minutes on stolen sessions, including the Pixel 10a reuse pattern. A token already issued is good for that window of access only. With MFA 2.0 as the only sign-in and enrollment method, a stolen session token is only that login. Once the token expires it cannot be reused, and the attacker cannot add a device. The primary win remains upstream: the proxied login never succeeds, so the enrollment race never starts.

Control Against live AiTM login Against later session replay
Password + push/OTP Relayable in real time N/A after session exists
Passkeys at login only Origin-bound at ceremony Does not revoke issued tokens
Device-bound phishing-proof MFA No transferable factor to relay Does not undo tokens already out
Session revoke + device disable Cleanup after theft Required residual response

Malware that steals cookies from a machine that already holds a legitimate session is a harder residual path. This wave’s documented entry was the cheaper one: coach the worker, proxy the login, keep the session. Closing that path forces attackers up the cost curve instead of letting a phone call finish MFA for them.

Passkeys harden origin binding at login

Passkeys and FIDO2-style WebAuthn credentials place a key pair on the device, keep the private key in hardware, and produce an origin-bound signature at authentication. Industry language correctly calls that phishing-proof at login: a reverse proxy on the wrong host cannot obtain a signature meant for the real enterprise identity provider. On a generic enterprise IdP that issues only hardware-bound signatures, the worker unlocks a local private key rather than typing a reusable secret the AiTM page can ship upstream.

That login hardening is real and necessary against the Com proxy pattern. It is still not the same as phishing-proof MFA 2.0 as the only sign-in and enrollment method. With MFA 2.0 as that only method, a stolen session token is only that login, and the attacker cannot add a device. For this Unit 42 wave the public record centers on proxied password-plus-MFA completion and post-auth persistence, not a documented Temporary Access Pass reset, so the decisive gap on the page is the relayable login itself.

Defenders who only compare “MFA on” versus “MFA off” miss the mechanism. The Com clusters completed MFA. They completed it through someone else’s site. Origin-bound device keys refuse that completion. The device-bound architecture that never ships a reusable factor is what makes the refusal stick across the environments workers actually use.

Key takeaways for defenders

Workforce identity will keep drawing voice coaches and tailored reverse proxies while transferable MFA remains the default. The fix is not a louder prompt. It is a login ceremony that cannot be completed on the attacker’s origin in the first place.

FAQ

Would phishing-proof device-bound MFA have stopped the Com-affiliated AiTM login?

Phishing-proof device-bound MFA would have stopped the Com-affiliated AiTM login path Unit 42 documented because that path depended on relaying a workforce password and a live MFA approval through a reverse proxy. With no transferable secret and an origin-bound device signature, the lookalike portal cannot finish a usable Microsoft 365 sign-in for the caller.

Can MFA 2.0 undo Graph mailbox theft after a session is already stolen?

MFA 2.0 cannot undo Graph mailbox, OneDrive, or SharePoint theft after a Microsoft 365 session is already in attacker hands. In the Com-affiliated campaigns, that data access used captured sessions and Python Graph-oriented scripts after authentication. Revoke, authenticator disablement, and session hunting are the residual controls once the token exists.

How do passkeys compare to phishing-proof MFA against this AiTM pattern?

Passkeys are phishing-proof at login because they produce an origin-bound hardware signature the Com-style reverse proxy cannot complete on a spoofed host. With MFA 2.0 as the only sign-in and enrollment method, a stolen session token is only that login: once it expires it cannot be reused, and the attacker cannot add a device.

What should defenders do about the Pixel 10a style rogue MFA enrollment?

Defenders should revoke the stolen session and remove unauthorized MFA devices. Unit 42 observed CL-CRI-1264 enrolling unauthorized MFA devices within minutes, including reuse of a device named Pixel 10a across accounts. With MFA 2.0 as the only sign-in and enrollment method, a stolen session token is only that login, and the attacker cannot add a device. Preventing the proxied login removes the session that made that enrollment race possible.

Does blocking AiTM vishing also stop the Helix browser-extension path?

Blocking AiTM vishing does not automatically stop the CL-CRI-1192 Helix browser-extension path Unit 42 described in the same Com-affiliated wave. That alternate method harvested authentication cookies and MFA tokens from a sideloaded extension, including via QR scans, which is endpoint and browser-control work rather than a login-ceremony fix alone.