darinor
Defense engineering

MFA Protects the Log-in, Not the Session

Replaying a stolen session cookie never triggers a prompt. Attackers stopped beating MFA — they log in with the session it already approved.

An Outpost24 benchmark published this month found credential exposure touching 72% of the 4,824 organizations it measured — and at 41% of them, at least one exposed login traced to a work email address, which usually means a work device was infected. Flare's 2026 infostealer report, built on 18.7 million logs from 2025, found 1.17 million carrying the worse combination: enterprise credentials and a live session cookie, harvested from the same machine.

A stolen password has to get past MFA. A stolen cookie doesn't. That difference is where most identity programs stop defending.

What the log contains

An infostealer doesn't break into an identity provider. It lands on a laptop — a cracked installer, a sketchy extension, and this year's favorite lure: fake AI app installers built specifically to harvest AI service accounts — and it empties whatever the browser already holds. Saved passwords. Autofill. Every session cookie, on disk or lifted straight out of browser memory while the page is still open.

A session cookie is what your identity provider hands out after you pass MFA. Keep-me-signed-in stretched it from an afternoon into a month, by design, because that's what users asked for. From the server's side, a replayed cookie is indistinguishable from you opening your own laptop.

The exposure concentrates where the access does. Flare found Microsoft Entra ID in 79% of enterprise identity logs in their 2026 report. One login fronts your tenant's mail, files, and cloud consoles. Which means one stolen cookie fronts all of it too, and the identity platform your team spent years hardening rides out the door in a browser profile.

MFA guards the moment of login. The session is what the attacker actually uses.

Replay is a marketplace

For years, infostealer logs were an intelligence footnote: passwords in bulk on marketplaces, mostly stale, mostly noise. Two things changed.

The targeting moved up the stack. In Flare's data, enterprise identities appeared in roughly 6% of infections at the start of 2024 and nearly 14% by the end of 2025 — while total infostealer volume declined. Fewer infections, each worth more: the trade shifted from stealing everything to stealing what opens doors.

And the sessions became a product. Okta's threat intelligence team published a September analysis of a dump released for free on Telegram in August, and catalogued 44,791 access tokens inside it — 1,843 of which were still unexpired the day the data went public. The buyers aren't manual, either: they replay stolen sessions through anti-detect browsers and automation frameworks that mimic the victim's device fingerprint, because "login from a new device" is exactly the check those tools were built to pass. The 2024 Snowflake intrusions were the password version of this economy. The current wave skips the hard part — there's no password challenge to fail.

The reset reflex

Someone reports suspicious access, or a monitoring platform surfaces an infostealer hit. The reflex is to reset the password. It feels decisive. It changes nothing the attacker needs.

Borja Rodriguez, Outpost24's head of threat intelligence: "An infostealer doesn't stop at one password. It can take session cookies, saved credentials, and access to multiple accounts." A password reset touches exactly one of those, and it isn't the one being used.

Resetting a password corrects a fact the attacker already copied. Revoking the session ends the access itself.

The sequence that actually works reads backwards from instinct:

  1. Revoke every active session for the affected account. Your identity provider has a per-user "sign out everywhere" switch; each SaaS console underneath it is its own hunt, and that's precisely why the steps belong in a runbook instead of being improvised under pressure.
  2. Then reset the password — once nothing the attacker holds outlives the change.
  3. Investigate the device, not just the account. A stolen credential with its sessions revoked is a datapoint. An active cookie in someone else's browser is an ongoing intrusion, and the machine it left behind is the source of the next one. The incident isn't over during cleanup, either — this is that principle applied to identity.

What actually binds a session

The session problem finally has a security answer, and it isn't another prompt. The pattern running through the real fixes is the same one from voice cloning: move the trust somewhere the attacker's channel can't reach. There, it was a number you saved before the call. Here, it's the device.

Device-bound sessions. Chrome's Device Bound Session Credentials (DBSC) cryptographically tie a session cookie to a hardware-protected key on the device it was created on — an exfiltrated cookie fails its check on replay instead of silently working. Google shipped it to Windows users in Chrome 146, with macOS coming; Google Workspace is rolling the enforcement out, and the platform spec is public at the W3C. This is the first mainstream mechanism that turns "your session" into "your session on your machine."

A login you can't phish. Passkeys and FIDO2 keys keep the credential out of the infostealer's reach in the first place — the factor never gets typed, stored, or replayable.

Sessions the size of the blast radius. Session lifetime is a control, not a user preference. Admin consoles get short lifetimes and prompt re-auth; the internal forum can keep weeks. Right-size per app, and write the sign-in frequency choices down — if every session in your tenant lives at "maximum convenience," that default is a decision someone made without deciding it.

Monitoring tuned for replay. Impossible travel, a session fingerprint that changes browsers, token use at hours the device never wakes up. Weaker than the other three, and the replay kits were built to defeat it. Still the only one of the four that finds the intrusion already in progress.

Before hardening any of it, look at what your identity provider issues today: decode one real session token with the JWT decoder and read the expiry your users' browsers are holding, then walk the auth flow itself through the OAuth/OIDC debugger before you change anything you can't observe.

What this cannot do

An honest version of this post has to say where it stops.

Device-bound sessions raise the floor, not the ceiling. DBSC only protects the applications that adopt it, on the browsers that implement it, on the platforms that shipped it — Windows first, macOS in flight. Every app in your tenant that never opted in still hands out bearer cookies. Coverage is a deployment matrix, and most organizations don't know theirs.

Passkeys bind the login, not the session — replay of a post-login token is untouched by them. They block the path to a session, which is half this problem, which is not all of it.

The blind spot is personal. Your managed fleet is audited; the laptop your CFO uses to check mail from the kitchen table isn't, and Outpost24's 41% isn't only happening on corporate machines. The infostealer log you never see is the one doing the damage.

And device fingerprints can be forged — the replay kits are commercial and good at it. Treat detection as a tripwire with a lag, not a wall.

This week

Revoke every active session on your highest-privilege admin account — actually now, it's one checkbox in your IdP's console. Then make the step survive its next use: "revoke all sessions" goes into the offboarding template next to "collect the laptop," and into the incident runbook immediately above "reset password."

Three minutes per account, no vendor invoice. Outpost24 found stolen credentials at 72% of the organizations it measured. The uncomfortable question isn't whether you're in that group. It's whether anyone at your place has ever once clicked the revoke button.