Security teams that only score “users with MFA enabled” would have marked Key Eater green while the campaign was already live. According to the Microsoft Security Blog on 5 November 2026, operators used the ARTEX agentic AI framework to search public GitHub for exposed Entra application client secrets, then completed service-principal authentications with those valid secrets without any interactive user MFA challenge. Key Eater is the late-2026 proof that workforce MFA, including phishing-proof device-bound login for people, does not sit in front of app-only client-secret authentication, so prevention has to kill long-lived workload secrets rather than add another human second factor.

Why Entra app client secrets never faced user MFA

ARTEX-driven secret hunting against public repositories began in August 2026. Operators fed the agentic tooling large volumes of queries for organization names, secret keywords, and cloud and AI service strings. The objective was industrial discovery of credentials already sitting in public git history, not a novel break of Entra cryptography.

Microsoft recorded successful service-principal authentications from 199.19.77.84 across multiple application IDs and tenants from about 28 October 2026. Microsoft stated there was no product vulnerability: possession of valid secrets by themselves enabled app-only access. Publicly exposed credentials remain usable until revoked or rotated, and removing the original disclosure alone does not remediate the exposure. Organizations that find secrets in public repositories must assume those secrets and associated applications are compromised.

Todyl’s overlapping telemetry, which Microsoft cross-referenced on the same infrastructure, put numbers on a monitored slice. First stolen-app access hit at 08:45 UTC on 1 November 2026. By 3 November, bulk downloads and sweeps crossed 15 client tenants. In that set, 45 app registrations were targeted, more than 33 authentications succeeded, 27 organizations were touched including MSP downstream clients, 81 files were downloaded, and about 2,279 mailbox items were accessed. Compromised apps requested Microsoft Graph and Key Vault tokens, pulled SharePoint and OneDrive files, read mailbox items, and enumerated Azure subscriptions, resource groups, permissions, and service configurations. Multitenant and vendor-owned application credentials authenticated into downstream customer tenants. Todyl also saw Azure DevOps user session-token replays, with Conditional Access later blocking some refresh-token replays.

One Microsoft-observed case collapsed the clock: an AI agent moved from a GitHub secret scan to Azure Key Vault exfiltration in five minutes. Individual victim company names are not public, and totals outside Todyl’s monitored set are unknown. The full attack-chain write-up lives on the companion site: Key Eater stole Entra app secrets via ARTEX.

Entra application client secrets are long-lived shared secrets bound to an app registration. Whoever presents the secret authenticates as the application. That ceremony is app-only by design. It does not open a workforce login page, does not issue a push, does not wait for an OTP, and does not run a FIDO assertion for a person.

Stage What was used Human MFA in the path?
Initial access Stolen Entra app client secret No, app-only sign-in
Lateral movement Multitenant or vendor app reuse No, secret reused across tenants
Data access Graph, Key Vault, session or refresh tokens No, post-sign-in token use

Microsoft’s adaptive probing detail underlines automation, not confused humans. Key Vault metadata enumeration failed with HTTP 400, then succeeded about 28 seconds later after the client switched to ARM API version 2023-07-01 and received HTTP 200. Target shopping lists focused on storage account keys, App Service settings, connection strings, publishing credentials, AI and ML connections, AKS credentials, and Foundry Tools-style AI service credentials. Compromised OAuth apps also drove Run Command on Azure VMs. None of that sequence required defeating a person at sign-in.

What phishing-proof people login still cannot cover

Device-bound, origin-bound authentication for people prevents remote completion of a human login when the factor is a hardware-held private key and nothing transferable is typed or approved. That model is the right answer for workforce phishing, helpdesk-coached fake logins, and AiTM kits that need a live human ceremony. Key Eater’s primary path never started that ceremony.

App-only client-secret authentication sits beside the human identity lifecycle, not inside it. Workload identity secret reuse across multitenant apps is still secret possession, not a phishable user factor. Session and refresh token replay against Graph, Key Vault, SharePoint, mail, and DevOps happens after a credential that already succeeded. Prevention, not detection still matters for people; it does not rotate a client secret that never asked a person anything.

Phishing-proof MFA 2.0 would not have blocked service-principal authentication with a valid stolen app secret. it would not have blocked service-principal authentication with a valid stolen app secret, would not have stopped multitenant secret reuse into customer directories, and would not have un-issued tokens already granted to a compromised app. What it still does for the same organizations is close the separate human surface so operators cannot also harvest employee sessions with a parallel phishing kit while defenders are busy rotating apps.

Passkeys and WebAuthn-style factors are origin-bound at human login: a key pair lives on the device, the private key stays in hardware, and the signature is origin-bound so a lookalike page cannot complete the ceremony. On a generic enterprise identity provider that issues only hardware-bound signatures for people, that design stops remote relay of a transferable user factor.

Passkeys do not authenticate an Entra application client secret. They do not sit in the app-only client credential flow. They do not make a multitenant vendor secret unusable after it appears on GitHub. Treating “we rolled out passkeys to staff” as coverage for service principals confuses two identity planes that share a cloud tenant and almost nothing else at authentication time.

Sources do not describe helpdesk recovery, Temporary Access Pass coaching, or phishable new-device enrollment as the primary Key Eater vector. The door was a static app secret already public. Enrollment-gap analysis still matters for workforce programs elsewhere: origin-bound login can be undone if registration or device onboarding still trusts email OTP, SMS, or push. MFA 2.0 is phishing-proof across registration, device onboarding, authorisation, authentication, and decommissioning for people precisely so those human lifecycle stages stop handing attackers a transferable factor.

That full-lifecycle human guarantee still leaves workload secrets as their own problem. Architecture that binds proof to a device for interactive users is not a second factor for a client secret that was never meant to challenge a human. Kill the secret class; a stronger people-login product does not retroactively cover it.

Workload controls that close the Key Eater gap

Engineering culture still treats client secrets like configuration trivia: sample apps, CI logs, forgotten forks, and “temporary” secrets that never rotate. Multitenant vendor apps multiply the blast radius when those secrets leak. Downstream customers inherit the vendor’s secret hygiene whether they staffed a security team for it or not. Measuring only interactive user MFA coverage will mis-score this entire class as unexplained cloud access.

Workload prevention that actually maps to this breach looks different from a stronger second factor on users:

Control Effect on Key Eater-class abuse
Eliminate long-lived client secrets Removes the reusable shared secret ARTEX harvested
Prefer managed identities or federated workload credentials No static secret in git history to scrape
Immediate rotate and revoke on public exposure Ends use after discovery, not after repo cleanup alone
Shrink multitenant app permissions Limits downstream tenant blast radius
Token revoke plus Conditional Access on replay Interrupts encore sessions after initial secret use

Microsoft’s own remediation language matches that table. Exposed secrets stay usable until revoked or rotated. Repo deletion without rotation is theater. Assume compromise of the app registration, every permission it held, and every tenant that trusted it. Short token lifetimes force re-authentication sooner; they do not invalidate a client secret still sitting in an attacker’s notes.

Key Eater’s core chain was not that SSO cookie story. It was workload secret possession. User-session detection bolted onto a broken app-secret inventory does not finish the job.

Five-minute agent loops from public secret to Key Vault will keep winning against shared secrets that never expire. Human phishing-proof authentication remains mandatory for employees, contractors, and admins. It is not a substitute for deleting the app-secret pattern that authenticated Key Eater without ever ringing a user’s phone.

FAQ

Would phishing-proof MFA 2.0 have stopped Key Eater?

Phishing-proof MFA 2.0 would not have stopped Key Eater’s primary path. Key Eater authenticated with stolen Entra application client secrets as service principals, which is app-only access that never challenges a user for MFA. Device-bound proof for people closes human login phishing; it does not invalidate a valid workload client secret already in attacker hands.

Why did user MFA never apply to Key Eater service-principal sign-ins?

User MFA never applied because service-principal authentications with application client secrets do not challenge a human. According to Microsoft’s 5 November 2026 Security Blog, valid secrets by themselves were enough for app-only access and no product vulnerability was required. Framing Key Eater as push fatigue or OTP relay misstates the mechanism Microsoft documented.

Do passkeys or FIDO2 stop stolen Entra app client secrets?

Passkeys and FIDO2 do not stop stolen Entra app client secrets. They are origin-bound controls for interactive human login, with origin-bound signatures and private keys that never leave the device. App-only client-secret authentication is a different credential type; preventing Key Eater-class abuse means removing or rotating those secrets and preferring non-secret workload identities.

What should replace long-lived Entra application client secrets after Key Eater?

Long-lived Entra application client secrets should be replaced with managed identities or federated workload credentials that do not place a reusable shared secret in source control. After any public exposure, organizations must rotate secrets, revoke issued tokens, and re-scope multitenant permissions, because Microsoft states exposed credentials remain usable until revoked or rotated even if the GitHub file is deleted.

How should CISOs measure identity risk differently after Key Eater?

CISOs should measure workload identity exposure separately from workforce MFA coverage after Key Eater. Todyl’s monitored slice alone saw 27 organizations, including MSP downstream clients, reached through compromised app registrations and follow-on token use. A green “users on MFA” dashboard can coexist with public client secrets that never present a human factor, so secret inventory, multitenant blast radius, and rotation speed need their own scoreboard.