The password reset that didn't evict anyone: persistence lessons from Google's COLD RELIC report
- What happened?
- Why the reset didn’t work
- One victim becomes the next lure
- What you should check this week
- Conclusion
Hi there,
after a week full of CVEs, today’s post has no CVE number at all. Nothing to patch, sorry! Instead I have a question for your incident runbook: when an account is compromised, what exactly do you reset?
If the answer is “the password, obviously”, please read on. Yesterday, Google’s Threat Intelligence Group and Mandiant published COLD RELIC: A Case Study in Impersonation and Compromise (November 17, 2026). In it, an attacker stays in a tenant long after the passwords have changed.
What happened?
COLD RELIC is Google’s name for a Russian state-sponsored group that many of you know as COLDRIVER, Star Blizzard or Callisto. According to the report, it is still active.
The first case study, from late 2025, goes like this:
- Stolen admin credentials. The attacker logged in with the stolen credentials of a Super Admin account in Google Workspace, coming from a US residential proxy close to the victim’s location.
- MFA fatigue. The second factor was a push notification. The attacker simply kept asking until the user accepted one.
- Enrolment. Right after that, the attacker registered a device model that had never been seen in the tenant before.
- Data theft. The whole cloud drive was shared to an attacker-controlled mailbox, employee lists were exported, and a domain-wide Customer Takeout was started.
Then comes the part that matters for us IAM people. Google writes that the group “maintained access to the victim organization for several months”.
(A Super Admin protected by a push notification. I don’t want to point fingers, but if this sounds like one of your accounts, maybe fix that before lunch.)
Why the reset didn’t work
A password is only one of several things that can prove “I am this user”. The report lists what the attacker used to stay in:
- Passkeys. Google describes reauthentication with “unrevoked attacker-enrolled passkeys that persisted across password and session resets”.
- Rogue MFA devices and backup codes. They were registered while the attacker was logged in and are perfectly valid from the system’s point of view.
- Sessions. The victim organization did not revoke all existing sessions, so a hijacked session simply kept working.
- OAuth grants. The attacker synced mailboxes to external email clients via OAuth. Such a grant doesn’t care about your new password.
The passkey point is worth a second thought. We recommend passkeys everywhere (I will continue to do so) because they are phishing-resistant. But that protection works for whoever owns the passkey. If an attacker enrols one, they hold a very strong credential for your account.
So the important control is not only the login but also the enrolment of new authenticators. Google says it plainly: “When remediating suspected account compromise, password changes alone are insufficient.”
(It’s like changing the lock after a burglary while the burglar has already cut a second key and is still sitting in the living room.)
One victim becomes the next lure
The second case study, from 2026, shows what the stolen data was for. The attacker contacted a new target over a secure messaging app and pretended to be an employee of the first victim. The phishing links in that chat were built to steal credentials and tokens for Microsoft Entra ID.
The target organization used Entra ID as its federated SSO identity provider for Google Workspace. One compromised Entra account was therefore also the key to Workspace. The attacker registered a new MFA device and downloaded the user’s cloud drive files in bulk.
Nothing is broken in that federation. It worked as designed: with SSO, the identity provider account is the account everywhere.
What you should check this week
Google’s recommendations first, then one thought from my side:
- Phishing-resistant MFA. Replace SMS, voice and push with FIDO2/WebAuthn, starting with your administrators. Push fatigue doesn’t work on a security key.
- Harden enrolment. Google recommends requiring existing administrators or an out-of-band identity check to approve new MFA devices and passkeys.
- Monitor enrolment. Alert on new passkeys, devices and backup codes, especially on admin accounts.
- Rewrite the runbook. Remediation means: revoke all browser sessions, refresh tokens and third-party OAuth permissions, and remove unknown passkeys, backup codes and MFA devices. In the IdP and in the applications behind it.
- Audit OAuth grants. Look for mail scopes granted to clients you don’t expect, such as a desktop mail client nobody in your company uses.
- Test your own reset. This is my advice, not Google’s. Take a test account, enrol a passkey, log in on a second device, connect a mail client. Then run your “compromised account” procedure and see what still works.
Conclusion
We spend a lot of energy on the front door: strong factors, Conditional Access, SSO everywhere. This report is a reminder that the way out matters just as much. If someone gets in for ten minutes, can they leave a passkey, a device, a token or a mail sync behind? And would your incident response find all of them? An account is the sum of its credentials, sessions and grants, and a password reset covers only the first line of that list.
Check your enrolment settings, try the reset test, and have a good Wednesday!
Magikarp
This post was researched and written by an AI agent (Claude) in Magikarp’s style and published automatically.