A managed law-firm laptop was already signed into Microsoft 365 when the spoofed IT video call began. MFA and Conditional Access had cleared on that compliant device. According to ReliaQuest Threat Research, Luna Moth (also tracked as Silent Ransom Group) operators then coached the employee to copy a live Microsoft Graph bearer access token from Microsoft 365 Copilot browser DevTools, paste it to an attacker-controlled site, and watch attackers replay it from their own infrastructure within about a minute. Thousands of files left through the Graph API in roughly an hour. No attacker sign-in event. No endpoint malware. Login MFA never got a second chance to run.

ReliaQuest maps Luna Moth’s Copilot token handoff

ReliaQuest Threat Research, authored by Joseph Keyes and published 12 November 2026, describes an October through early November 2026 wave against law firms the group had already pressured in earlier extortion rounds. The lure opened with a blank-subject free-mailbox message carrying a deliberately broken PDF. A spoofed IT-support call followed, often to a personal number, offering to “fix” the bad attachment on a video session with screen share.

On that call the operator directed the employee to open Microsoft 365 Copilot in the browser on the managed device, open developer tools, find the Microsoft Graph bearer on the Copilot-to-Graph request, and paste the string into a public text-sharing site the attackers controlled. ReliaQuest reports replay from attacker infrastructure within about a minute and bulk Graph API download of thousands of files in roughly an hour. Public reporting has not named the firms, has not published victim counts, and has not named the video-call platform beyond noting it is typically permitted software.

Earlier Luna Moth work against law firms, documented by the same ReliaQuest spotlight since at least mid-2023, used IT-support vishing plus remote-management tools and manual downloads for extortion without encryption. The newer path keeps the social script and swaps the exfil channel for a Graph bearer the employee exports after a normal session already exists. The full attack-chain narrative, lure detail, and timeline live on the companion breakdown of the live Graph token handoff.

After sign-in, the Graph bearer became the prize

The Microsoft Graph access token was issued after the employee completed sign-in on a managed compliant device. MFA and Conditional Access were satisfied at that moment. ReliaQuest states the stolen token passes Conditional Access because the device check runs at sign-in, not when the token is later used. Once issued, that bearer string is portable proof of access until it expires or another control cuts it off.

Default Microsoft access-token lifetime sits around 60 to 90 minutes. Continuous Access Evaluation can stretch supported tokens much longer or revoke them early when conditions change, such as location or account disablement. Without CAE and strict location enforcement, a token stolen on a trusted laptop keeps working from attacker servers until the clock runs out. Conditional Access does not re-evaluate every Graph call the way a fresh interactive login would.

Control At issuance On later Graph use
Compliant-device MFA Already satisfied Not re-run
Conditional Access device check Ran at sign-in No per-call re-check
Sign-in logs Employee login only No attacker login event
Endpoint malware defenses Nothing on disk Pure off-box bearer replay

The social engineering target was session material sitting in the browser, not a password field or a push prompt. ReliaQuest notes the technique shares a core idea with device-code phishing: the victim takes part in a legitimate Microsoft process that produces a valid token instead of typing a password for the attacker. Here the token is issued to the victim’s device first, then walked out of band through DevTools and a paste site.

Attackers never needed malware on the laptop. They never needed a second authentication ceremony. They needed one coached copy-paste while screen share made the steps feel like ordinary IT support. For a law firm, Graph scopes on that identity can reach client files, mail-adjacent content, and anything else the account’s permissions already allowed. The pressure model stays pure data extortion, including second-round theft from firms that earlier appeared on the group’s leak site.

Shortening token lifetime only shrinks the window. It does not stop a one-minute handoff followed by an hour of Graph pulls. If lasting access is registered during a window of control, lifetime math stops mattering for later revisits. Defenders who only hunt failed logins or unknown devices will miss a path that never creates either signal.

Device-bound login cannot recall an exported bearer

Phishing-proof device-bound authentication stops remote kits from completing a fake or proxied workforce login. That is a different problem than the one Luna Moth ran here. The employee had already finished a legitimate compliant-device sign-in. The private key ceremony, if any, had already succeeded. What left the building was a bearer string the browser already held for Graph calls.

No login control, including device-bound keys that never leave hardware, can un-issue a token the user is talked into exporting after authentication. The call did not harvest a password, OTP, push approval, or recovery secret. It harvested post-sign-in API material. MFA 2.0 does not monitor sessions, score behaviour, or revoke Graph bearers after they exist. That control does not have that capability.

What still matters upstream is closing every path that issues tokens to attackers in the first place: AiTM reverse proxies, device-code coaching, helpdesk re-enrollment with transferable factors, and spoofed login pages. Those remain credential-phase failures where origin-bound, device-bound signatures remove the transferable factor before a session exists. Luna Moth’s observed path skipped that surface. The session was already good.

Containment for this class is revoke-and-bind work. Continuous Access Evaluation with hard location rules can make a stolen bearer stop working outside the victim network, which is exactly the control ReliaQuest recommends. Fast session and token revoke, Graph bulk-download detection, and tight visibility on anomalous API volume matter more than another login prompt. Training must treat “open DevTools and paste this string” as an incident, not an IT chore. Hang up. Call a known internal number. Bearer material pasted into external sites under screen share is how a live Graph token leaves the device.

Architectural designs that avoid long-lived, freely replayable bearers for high-value cloud APIs reduce blast radius when social engineering still wins a paste. Per-app, fresh device-bound proof with no shared replayable federation cookie shrinks the skeleton-key problem classic SSO creates across many apps. The Prevention, Not Detection boundary still draws a hard line: stop phishable issuance when you can; once a bearer is live and pasted off-box, detection and revoke carry the load.

Stage What attackers held Login MFA effect
Initial handoff Live Graph bearer after compliant sign-in Already satisfied; no second check
Graph bulk download Replay from attacker infrastructure No interactive auth barrier
Extortion pressure Thousands of files, no encryptor Post-auth data loss

Origin-bound passkeys still leave post-login paste risk

Passkeys and FIDO2-style WebAuthn put a key pair on the device, keep the private key in hardware, and produce an origin-bound signature at login. That design is origin-bound authentication: a reverse-proxy fake origin cannot complete the ceremony with a typed secret. On a generic enterprise identity provider that issues only hardware-bound signatures, the login act itself hardens sharply against classic credential phishing.

That resistance still applies only to authentication. It does not stop an employee who is already signed in from copying a Graph bearer out of DevTools. Luna Moth’s call never needed to defeat the login ceremony. The ceremony had already succeeded on the real compliant device. Passkeys would not have made the pasted string worthless by themselves. Token binding, CAE, short practical lifetimes paired with revoke discipline, and human process controls against out-of-band token export sit in a different layer.

Recovery flows were not how Luna Moth got the token

Public reporting does not describe helpdesk password resets, Temporary Access Pass issuance, or new-device enrollment in this wave. The employee used an already-working managed session. Enrollment and recovery were not the door ReliaQuest documents here.

That still matters for identity architecture elsewhere. Many real breaches do abuse phishable recovery: SMS codes read aloud, email OTPs, push approvals on a new device, or support staff re-enrolling an attacker. origin-bound passkeys harden login only. Phishing-proof MFA 2.0 removes phishable factors across registration, device onboarding, authorisation, authentication, and decommissioning so recovery cannot reintroduce a transferable secret. Luna Moth simply did not need that path. The lesson from this campaign is post-auth bearer handling, not a recovery redesign for this specific handoff.

Key takeaways for defenders

Client confidentiality, notification duty, and reputation risk sit on top of the technical loss. Law firms hold privileged material by design. A quiet Graph bulk pull from attacker infrastructure is a different operational nightmare than ransomware that trips every alarm on the floor. Stronger login cryptography remains essential against the campaigns that still steal sessions at the form. For Luna Moth’s observed Graph paste path, the practical fix is making stolen bearers die quickly and teaching staff that no real IT team will ever ask them to export one.

FAQ

Would phishing-proof MFA 2.0 have stopped Luna Moth’s Graph token theft?

No. Phishing-proof MFA 2.0 would not have stopped the Luna Moth Graph token theft path ReliaQuest describes, because the Microsoft Graph bearer was issued after a legitimate compliant-device sign-in that already satisfied MFA and Conditional Access. The attackers coached an out-of-band paste of that live token and replayed it. Login controls do not recall a bearer the user exports after authentication.

Why did Conditional Access allow Luna Moth’s off-box Graph downloads?

Conditional Access allowed Luna Moth’s off-box Graph downloads because, as ReliaQuest states, the device check runs at sign-in and does not re-check the token every time it is used. A bearer stolen after a compliant login remains valid from other locations until expiry unless Continuous Access Evaluation and strict location enforcement cut it off.

How is Luna Moth’s Graph handoff similar to device-code phishing?

Luna Moth’s Graph handoff is similar to device-code phishing in that the victim completes a legitimate Microsoft process that yields a valid token rather than typing a password for the attacker. ReliaQuest draws that parallel explicitly. The difference is issuance target: device-code flows often land the token on attacker-controlled infrastructure, while this wave issued the token to the employee’s compliant device first and then walked it out through DevTools and a paste site.

What should law firms deploy if login MFA already passed in this scenario?

Law firms facing Luna Moth-style Graph bearer paste should deploy Continuous Access Evaluation with hard location rules, rapid token and session revoke runbooks, Graph bulk-download detection, and explicit staff rules against pasting DevTools secrets under any remote support call. ReliaQuest highlights CAE with strict location enforcement so a stolen token stops working outside the victim’s network. Those controls address post-issuance replay. They are not substitutes for phishing-proof login on other campaigns that still attack the sign-in form.

Did endpoint malware play a role in the observed Luna Moth token wave?

No. ReliaQuest reports no endpoint malware and no attacker sign-in events in the observed Luna Moth Graph token activity. Operators replayed a bearer the employee pasted from a real Copilot browser session and pulled thousands of files through the Graph API from their own infrastructure in roughly an hour.