Google says COLD RELIC used one US victim's staff data to pose as its employee and breach a second
Google Threat Intelligence Group and Mandiant say the Russian group COLD RELIC used a staff directory and identities stolen from one US organisation in late 2025 to pose as its employee and approach a second in 2026. Neither victim, no day and no full count is stated.
By Parminder Kumar Sharma · LinkedIn · · 19 min read
The checkable link between the two victims is one URL parameter, and the page prints no day for it
Google Threat Intelligence Group (GTIG) and Mandiant's page on the Google Cloud Blog, dated 17 November 2026 in its own page data, says that during a 2026 intrusion at a US-based organisation, "COLD RELIC initiated contact over an encrypted messaging app while impersonating an employee from the first victim organization". The page does not print that conversation. Its Figure 2 is captioned "Redacted, LLM-generated depiction of observed threat actor social engineering engagement", so the chat shown is an illustration, not a transcript. Besides that conversation, the page gives one other item that links the two victims, and it is the checkable one: GTIG "correlated a COLD RELIC redirector URL" that had been "submitted to VirusTotal in June 2026, whose query parameter referenced the first victim organization". The page prints no day for that submission. From 30 June to the page date of 17 November is 140 days and from 1 June it is 169 (derived: the two ends of the month), so the submission was between 140 and 169 days before the page appeared.
What that does not establish. It does not establish who either organisation is, what sector either is in, how many organisations the group has compromised in this chain, how the first victim's Super Admin credentials were stolen, how the messaging app contact was started, or whether data from either victim has been published or used elsewhere. The page names no victim and gives no sector. It also does not say that the first victim's data was in the group's hands in June 2026: it says only that the URL was submitted then and that its parameter referenced the first organisation. This site names no victim and guesses no sector.
Whose account this is. The page is Google's account of its own incident response work. Google owns GTIG and Mandiant, sells security products and services (the page prints seven rule names from Google SecOps), provides the Workspace environment described in the first case study, and offers the Government-backed attack alerts the page mentions; each has an interest in the work being read as effective. This site read no independent source and no statement from either victim. COLD RELIC, COLDRIVER, UNC4057, Star Blizzard, Callisto and Gossamer Bear are group designations as the page prints them, not people. An earlier brief on this site, Nine of 13 Star Blizzard campaigns are event invitations, and the invitation usually has no attachment, covers Microsoft's account of campaigns it tracks under the Star Blizzard name; this site does not compare Microsoft's and Google's descriptions of the group here.
What happened, in the order the page gives it
Case study 1, late 2025. The page says GTIG observed COLD RELIC operators compromise the cloud productivity environment of a US-based organisation. For initial access the group used a residential proxy network in the US "located in close geographical proximity to the victim", and authenticated with the stolen credentials of a Super Admin account together with an MFA push notification fatigue attack. After the victim user accepted an MFA request, the group registered a device model not previously seen in the tenant, and kept access with passkeys, MFA backup codes and rogue MFA devices.
With Super Admin privileges the group, the page says, took the whole organisation's cloud storage drive by mass sharing files to an account it operated and by downloading files directly. In the Admin Console it "queried Admin security investigation and email audit logs to check defensive visibility" and "viewed, modified, and deleted Alert Center notifications to remove forensic evidence of the initial intrusion". It exported full employee lists and started a domain wide Customer Takeout export. The page says the group then kept access "for several months" (days not stated) and did three more things it lists: synced an email inbox by OAuth to external email clients, signed in again with attacker-enrolled passkeys that had not been revoked, and opened cloud drive documents about credentials, email archives and personnel information. When the group was removed, the page says, it used email accounts impersonating employees of the victim: "These impersonated accounts were used to target both internal employees and associates within the same professional circles with renewed social engineering attempts".
Case study 2, 2026. The page says the second organisation, also in the US, used Microsoft Entra ID as its federated single sign-on identity provider for Google Workspace, and that the group compromised an Entra ID user account there. To get in, it impersonated an employee of the first victim to set up a pretext for conversation on a secure messaging app, then shared phishing links in the course of the conversation. The links were "designed to steal credentials and authentication tokens for Entra ID" and redirected the victim to document lures relevant to the conversation. After the stolen credentials worked, the group registered and used a new MFA device in the single sign-on system to keep access. The page's own summary of the result is that the group "COLD RELIC was able to obtain credentials to the organization and perform bulk download of files accessible in the compromised user's cloud drive".
The page adds that, after the compromise, "the encrypted messaging application used for the initial approach recorded that the threat actor changed their profile name to a completely new person", and that GTIG believes the new persona was likely used against other organisations in the same professional circles. It says the compromised-site first link is a hallmark of the group: a compromised website carries the first link and redirects to the real phishing domain, which hides that domain and lets the group fingerprint a visitor before deciding whether to redirect.
How firmly the page states each point about the chain, in the page's own words. The pattern in the right column is this site's reading of the wording.
| Point | The page's wording | Firm, hedged or attributed |
|---|---|---|
| The first victim's data was used against the second | "directly weaponized data, the corporate staff directory, and identities stolen" (introduction) | Firm, in the introduction |
| Contact and relationship context was used | "likely leveraged contact and relationship context from the first organization to target the second" (threat detail) | Hedged: "likely", in the section after the introduction |
| The group sought data for intelligence | "likely to fulfill broader intelligence requirements" (late 2025 download) | Hedged: "likely" |
| Where the group sits | "US Department of Justice (DOJ) has assessed" it operates within Russia's Federal Security Service | Attributed to the DOJ. The page prints no confidence statement of its own |
| What the group does after a compromise | "primary objective post-compromise is mass collection and download of files and email archives" | Firm |
| The new persona was used against others | "GTIG believes this new persona was likely used" | Hedged twice: "believes" and "likely" |
The dates the page prints, and the ones it does not
The page prints five time references and no day. The first intrusion is "late-2025" (the page writes both "late 2025" and "late-2025"). Access after the theft lasted "several months". The redirector URL was submitted to VirusTotal in June 2026. The second intrusion is dated "2026". New phishing operations are dated "as recently as November 2026". The page itself is dated 17 November 2026.
This site derives one span only, because only one has two printed ends: from the June 2026 VirusTotal submission to the page date, 140 to 169 days (30 June to 17 November, and 1 June to 17 November), about 5 calendar months by month name. It does not derive how long the first intrusion lasted, how long "several months" was, or the gap between the two intrusions: the page gives no month for the first and no month for the second. It also does not say whether the second intrusion began before or after the June submission. Whether the November 2026 phishing operations involve either victim is not stated either: the page says only that GTIG has "observed evidence of new phishing operations as recently as November 2026".
Persistence and anti-forensics, and what the page prints against each step
The page lists four ways the group kept or regained access in the first case study (a count made by this site, derived): rogue MFA devices, passkeys, MFA backup codes, and an email inbox synced by OAuth to external email clients. It says the passkeys enrolled by the attacker "persisted across password and session resets and bypassed device registration controls". In a separate paragraph it says that in at least one case a victim organisation noticed the intrusion and reset passwords for the affected accounts, but because the group had hijacked sessions and the organisation did not also revoke all existing sessions, the group kept access. The page does not say that these two accounts describe the same victim, so this site does not join them.
Anti-forensics in the victim's own console. The page says the group queried audit logs "to check defensive visibility" and viewed, modified and deleted Alert Center notifications. That means a defender who relies on alerts inside the same tenant is relying on a surface the page says an intruder with Super Admin rights changed. The page's answer is to alert on the deleting event; it gives no way to restore a deleted alert or to keep a copy outside the tenant, so this site draws that row amber.
What the headline version of the story leaves out
A headline wording noted by this site, and not read at source, summarises the story as stolen Google Workspace data used to breach a second US organisation. This site read no relay and cites none; it checks that wording against the page. The page says the first victim's data, specifically the staff directory and identities, was used to construct the pretext. It says the second victim was entered through a Microsoft Entra ID user account after a messaging app conversation and phishing links that stole Entra ID credentials and tokens. Google Workspace figures in the page as the product behind the first victim's tenant and as the product the second victim signs in to through Entra ID. Nothing in the page says Google Workspace, Entra ID or any other product was breached, had a flaw, or was the route in for the second intrusion.
The headline readings tested against the page. Stated and not stated are the page's, as this site read it.
| Reading | What the page prints | What it does not print |
|---|---|---|
| Workspace data was used to breach the second victim | A staff directory and identities from the first victim, likely with contact and relationship context | That the second victim's entry used any Workspace credential. The entry is an Entra ID user account |
| The second victim was hacked through Google | Entra ID is its federated login for Google Workspace; the user was phished and the group registered a new MFA device | Any flaw in Google or Microsoft products. Which cloud drive was downloaded in case study 2: the page says "cloud drive" |
| A transcript of the pretext exists | Figure 2, a redacted "LLM-generated depiction" | A transcript. The chat shown is an illustration |
| The group has hit many organisations through this chain | "a series of linked intrusions" over "the past year"; GTIG has "tracked and disrupted" its phishing campaigns since 2015 | A count. The page describes two case studies, not a total, and gives no success rate |
| The group is part of Russia's Federal Security Service | The DOJ "has assessed" it operates within the FSB | A confidence statement from GTIG on its own identification of the group's sponsor |
| The group deploys malware in these intrusions | The group is "known for" phishing and malware campaigns in general | Any malware, file hash or tool in either case study |
Where the page differs from itself
These are differences inside one page. This site reports them and resolves none.
- Firm against hedged. The introduction says the group "directly weaponized" the staff directory and identities. The threat detail says the group "likely leveraged" contact and relationship context. The first is about specific data, the second about context; the page does not say how the two statements relate.
- Objective against purpose. The post-compromise section states the group's "primary objective" as mass collection and download, with no hedge. The earlier threat detail says the late 2025 download was "likely to fulfill broader intelligence requirements". The page hedges the purpose of the collection, not the collection.
- A series against two. The page says Mandiant responded to "a series of linked intrusions" over the past year and describes two. It does not give a count of the series.
- Two statements about resets. Passkeys "persisted across password and session resets" (case study 1 list), and, in another case, sessions were not revoked after a password reset (hardening section). The two accounts may or may not be the same victim.
- Two spellings of one event.
CUSTOMER_TAKEOUT_CREATEDin the narrative, "Customer Takeout Created" in the Detections table. - An illustration for a conversation. Figure 2 is an LLM-generated depiction, per its own caption, not a transcript.
- Dates. "Late 2025" for the first intrusion, "2026" for the second and June 2026 for the VirusTotal submission. The page does not say the first victim's data was in the group's hands in June.
- Attribution. The statement that the group is state-sponsored and Russian is the page's own; the placement within the Federal Security Service rests on a Department of Justice assessment that the page links. No GTIG confidence statement for that placement is printed. This site did not read the linked DOJ page.
What to do, in order
What the page recommends, once. Identity and MFA: enforce phishing-resistant MFA (FIDO2 or WebAuthn) and replace SMS, voice and push with hardware security keys; restrict and audit passkey registrations; require existing administrators or out-of-band identity verification to approve new MFA devices; consider the Advanced Protection Program. Sessions and response: in a compromise, revoke all active sessions, refresh tokens and authorised third-party OAuth app permissions across Google Workspace and Microsoft Entra ID and remove unauthorised passkeys, backup codes and MFA devices, because "password changes alone are insufficient"; audit OAuth grants. Network: enrich sign-in logs with residential proxy and VPN reputation, alert on residential ASN shifts for privileged accounts, restrict administrative access from consumer ISPs, VPNs and known residential proxy ranges, and apply cloud storage data loss prevention rules. The page's Detections table maps eleven activities to alerts (derived: count of rows). The items below are this site's suggestions, in the order worth doing, and each says where it rests on the page.
Take this with you
This site's suggestions, from the page's account
- Administrators of Google Workspace and Entra ID: review the events the page lists for your tenant (Admin audit events, passkey and device registration, MFA backup code sign-ins, new OAuth grants with broad mail scopes, Customer Takeout, Alert Center deletions, user list downloads, investigation queries) and decide who is alerted when each fires.
- Treat a message on a messaging app from someone you know as unverified until it is confirmed on another channel you already trust. The page says the second intrusion began with an impersonated employee on such an app.
- Check what your Takeout setting covers before relying on it. The page says administrators can disable Google Takeout for users; it does not say that stops a Super Admin running a Customer Takeout export, so alert on that event as well.
- After removing an intruder, revoke registered passkeys, backup codes, MFA devices, sessions, refresh tokens and OAuth grants, not only passwords. The page says passkeys persisted across password and session resets.
- Keep copies of alerts and audit logs outside the tenant. The page says the group queried audit logs and viewed, modified and deleted Alert Center notifications, and it gives no way to restore a deleted alert.
- After any compromise, assume contacts and staff lists were taken: warn partners and associates from your own domain, on a channel the intruder does not control. The page says impersonated accounts were used to target associates after the group was removed.
- Before blocking whole consumer ISP ranges, test the effect on ordinary users: the page prints two example ranges and does not name who runs them, and this site did not look them up.
- If you read the headline as a flaw in Google Workspace, re-read the page: it describes stolen credentials, MFA fatigue, a phishing link and a federated login, and states no product flaw.
What the page prints
The page prints no file hashes, no malware names, no tool names for the group and no real domains: its one example host is a placeholder. What it prints that a defender can use is below, copied from the page, each item attributed in its caption. This site defanged the placeholder host and added nothing else.
The two example URL patterns. The page's Figure 1 caption says they are sample URLs "incorporating targeted individuals' names or lure topics" and that this pattern is usual but "not always the case". The parameter value is set by the attacker and, the page says, often shows a relationship to a previous victim or what the lure is. The host below is the page's placeholder, defanged here.
hxxps://compromised[.]example[.]com/some_path/some_file.php?file=<previous_victim_name>
hxxps://compromised[.]example[.]com/some_path/some_file.php?ref=<lure_name>
Log events. The narrative names one event; the Detections table names the others, each with the log it sits in. Hunt for them in your own tenant; the page gives no thresholds.
Google Workspace Admin audit log (narrative):
CUSTOMER_TAKEOUT_CREATED
GWS Admin logs (Detections table):
Customer Takeout Created
Batch Delete Operation Performed
User List Download
Investigation Query
GWS Drive logs:
Download
User Sharing Permissions Change
Visibility changed to shared_externally
Item search performed
GWS OAuth logs:
Authorize (broad scopes, for example https://mail.google.com/, to unapproved clients)
GWS events (log not named in the row):
Passkey enrolled
Device Registered
GWS User logs:
Login Challenge Type is Backup code
Search terms, clients and ranges the page gives as examples. The page says to watch Drive searches for terms about archives or credentials, to flag unexpected native desktop email clients or external mail services, and to restrict administrative access from consumer ISPs and known residential proxy ranges. The two ASN numbers are the page's examples; it names no operator.
Drive search terms to monitor: .pst, .bak, password, cred
DLP sensitive terms (cross-cloud): passwords, roster, .pst
Unapproved clients and services: Thunderbird (native desktop client), Proton Mail (external mail service)
Residential proxy ASN ranges (examples): AS5650, AS20115
The mapping and the rules. The Detections table maps each observed activity to a MITRE ATT&CK technique. Google SecOps customers get rules for the same activity under the Mandiant Frontline Threats rule pack; the page lists seven rule names. Both lists are Google's; the rules are a Google product.
MFA fatigue: T1621
OAuth abuse and mail sync: T1114.002, T1528
Automated data exfiltration: T1119
Bulk download: T1530
External sharing: T1567.002
Device and passkey enrolment: T1098.005
MFA backup code use: T1078.004
Alert deletion: T1562.008
User list export: T1087.004
Admin log querying: T1082
Sensitive file search: T1083, T1552.001
Google Workspace Successful MFA Fatigue Attack
Google Workspace Customer Takeout Initiated
Google Workspace Alert Center Alert Deletion
Google Workspace Directory User List Download
Google Workspace Drive Bulk External or Freemail File Sharing
Google Workspace Device or Passkey Enrollment Following MFA Fatigue
Google Workspace OAuth Broad Mail Scope Granted to Unapproved Client
Persistence (case study 1 and the post-compromise section):
rogue MFA devices
passkeys
MFA backup codes
email inbox sync via OAuth to external email clients
Dates as printed (no day for any):
late 2025 first intrusion
several months access kept after the theft
June 2026 redirector URL submitted to VirusTotal
2026 second intrusion
November 2026 evidence of new phishing operations (as recently as)
17 November 2026 date of the page
The question the page leaves open
The page's controls are about identity: hardware keys, approved MFA devices, revoked sessions, alerts on admin events. The second intrusion started with none of those: it started with a message that looked as if it came from someone in the same professional circle. The page's only advice that touches that moment is out-of-band identity verification, and it applies to approving new MFA devices, not to a chat on a messaging app. If an attacker holds your staff directory, how would a colleague at a partner organisation know the message is not from you?
Sources
- PrimaryCOLD RELIC: A Case Study in Impersonation and Compromise, page dated 17 November 2026: the introduction, threat detail, two case studies, post-compromise activity, remediation and hardening and the detections table. Saved and read in full by this site on 18 November 2026. Vendor's account of its own workGoogle Threat Intelligence Group and Mandiant (Google Cloud Blog)accessed 2026-11-18


