Operation Cymbal: Cold Room
Overview
Challenge: Operation Cymbal: Cold Room Platform: HackTheBox Difficulty: Easy Type: Blue Team / Threat Hunting
Three independent incidents at the fictional company Cymbal Investments, spread over about 8 days of logs. The task is to find patient zero, follow the attacker through the network and identify the final objective, using Google SecOps (SIEM search, UDM, raw logs), Curated Detections and Google Threat Intelligence (GTI) reports as reference.
Incidents
- PowerShell Download And Execute: drive-by download, process hollowing, Kerberoasting, data staging
- SHINYHUNTERS Identity Compromise: phishing, rogue MFA device, OAuth app abuse, database exfiltration
- AI-Enabled Malware: malicious VS Code extension that uses the Gemini API to generate code
Needed Tools
- Google SecOps (SIEM, UDM search, raw log view)
- Google Threat Intelligence (GTI) reports as TTP reference
General Approach
- read the briefing and write down the incident and rough time frame before searching
- widen the time window immediately (the incidents span several days)
- find patient zero first, do not start at “actions on objectives”
- click an entity and pivot from there instead of typing a new query
- verify every suspicion: is this normal maintenance or really an attack?
Incident 1: PowerShell Download And Execute
Kill Chain
flowchart TD
A["Drive-by download via marketingtrends-insider.com"] --> B["ChromeSetup.zip -> ChromeSetup.js"]
B --> C["wscript.exe runs ChromeSetup.js (PID 4192), parent: explorer.exe"]
C --> D["powershell.exe (PID 6124) downloads d27e15eb.exe from cdn-updater-distribution.net"]
D --> E["Process hollowing into svchost.exe (PID 624)"]
E --> F["Camouflage: GoogleUpdate.exe started as scheduled task"]
D --> G["Domain recon: net.exe Domain Admins + nltest.exe DC list"]
G --> H["Kerberoasting against CYM-DC-01, SPN MSSQLSvc/CYM-DB-01.cymbal.corp:1433"]
H --> I["Hash of svc_backup cracked offline"]
I --> J["NTLM logon as svc_backup on CYM-FS-01 (WinRM)"]
J --> K["Q3_Audits folder searched (T1083)"]
K --> L["Financial_Audit_2026.zip staged in Temp (T1074)"]
Step 1: The alert
The case starts with the curated detection “PowerShell Download And Execute” on CYM-DEN-W10-042 for the user MARCUS.VANCE:
The URL pretends to be an image (.png) but the file is saved as .exe, a classic masquerading trick. The pattern (compromised website, fake browser update, JS dropper, hidden PowerShell) matches the GTI malware profile of FAKEUPDATES.
The web proxy logs show the referer of the download.
Step 2: Initial execution
The dropper was started by a manual double click in the Downloads folder (parent process is explorer.exe, so nothing automated):
wscript.exe (PID 4192) runs ChromeSetup.js. Note: this PID is not a hollowed svchost.
Step 3: Domain recon
The same PowerShell session (PID 6124) that downloaded the payload starts two discovery commands:
First the Domain Admins are listed, two minutes later all domain controllers. This explains why the attacker later targets CYM-DC-01: they asked for it first.
Step 4: Process hollowing
The case summary named two PIDs (4192 and 2048) for the hollowed svchost. Both were wrong:
- 4192 belongs to
wscript.exe(see step 2) - 2048 is only the PID of the logging service (Sysmon) itself
The real hit is a long-running svchost.exe (user SYSTEM, 3-digit PID) that suddenly starts a GoogleUpdate.exe as a scheduled task for camouflage:
Step 5: Kerberoasting
On the domain controller CYM-DC-01, Windows Event 4769 (Kerberos Service Ticket Request) is the one to look for. This comes from the Security-Auditing provider on the DC, not from Sysmon on the client.
0x17 is RC4-HMAC. Modern domains use AES by default, so a requested RC4 ticket is a downgrade: RC4 hashes can be cracked offline much faster.
To find more events for the same account, search for the SID:
Step 6: Identity switch
A successful logon on the file server CYM-FS-01, coming from the original workstation:
The SID is identical to the ServiceSid of the Kerberos ticket. This proves that the cracked hash of svc_backup is now used for a direct NTLM logon.
Step 7: Data staging
First the target folder is searched. The parent wsmprovhost.exe shows that the session runs over WinRM (T1021.006):
13 minutes later PowerShell creates an archive in the temp directory (T1074, Data Staged):
Dead end: LINMINER-01
At some point the case also showed an unrelated host LINMINER-01 (GCP, cryptomining, HANGMAN.V2 / ROCKHATCH / CAKETEARS). The case audit log explains why: the SOAR alert grouping rule grouped by shared entities, among them the file explorer.exe, which exists on every Windows host. In addition, the mining alert was never reviewed by a human, it was closed automatically as “Lab test, NOT_MALICIOUS” by a demo job. This host is a separate incident and not part of the chain.
Incident 2: SHINYHUNTERS Identity Compromise
Kill Chain
flowchart TD
A["Phishing: sso-cymbal-portal.com, fake internal SSO page"] --> B["elena.rostova submits credentials via POST (/api/v1/auth/token)"]
B --> C["Attacker logs in with stolen creds, source: NordVPN/Cloudflare proxy"]
C --> D["Rogue MFA device registered as new factor (Okta)"]
D --> E["OAuth app 'tooglebox-recall' authorized"]
E --> F["Okta security mail 'A new MFA factor was added' permanently deleted"]
D --> G["Cymbal_Q3_Database_Architecture_and_Service_Accounts.pdf downloaded from Google Drive"]
G --> H["Credentials for svc_dbadmin found in the document"]
H --> I["Logon as svc_dbadmin from CYM-DATA-WS-007 to CYM-DB-01 (WinRM)"]
I --> J["cymbal_cust_q3.bak compressed with zstd to cymbal_q3_exfil.tar.zst"]
J --> K["Extortion note README-IF-YOU-SEE-THIS-YOUVE-BEEN-HACKED.TXT dropped"]
K --> L["Exfiltration via POST to 142.11.200.188:8888"]
The GTI report on the ShinyHunters brand (UNC6240) was the reference: vishing/phishing for initial access, OAuth abuse and exfiltration plus extortion.
Decoys
Two phishing domains had the maximum risk score (pagerisk=100):
| User | Fake domain | Date |
|---|---|---|
emil-palmer |
signin.office365x24.com |
29.10. |
warren-morris |
signin.accounts-gooqle.com |
17.10. |
Both are noise. A third lead (disabled 2FA enforcement by marta@mydomain.org, January) belongs to a different demo dataset: wrong domain and wrong date.
The real hit had a much lower score (pagerisk=25) because the domain imitates the company’s own SSO instead of a generic brand. Do not follow the loudest alert, follow the one that fits the story.
Useful query for credential POSTs:
Step 1: Phishing
The target is an API token endpoint, not a simple login form, which points to token theft.
Step 2: Account takeover
Seven minutes later the attacker logs in and registers a new MFA factor (Okta already rates it as WARN):
The Mac is the device of the attacker, not of Elena. This is the actual “initial identity compromise”.
Step 3: Covering tracks
Every MFA change triggers a security mail. A newly authorized third-party OAuth app deletes it before Elena can see it. The key was the log category applicationName: "gmail", found via the UDM event type EMAIL_TRANSACTION instead of free text keywords:
The app is called tooglebox-recall, a misspelling of the “ToggleBox Recall” tool from the GTI report. That is why searching for the original name found nothing.
Step 4: Sensitive document
With the taken-over account the attacker downloads a document from Google Drive. A file that documents database architecture and service accounts is exactly where credentials end up.
Step 5: Access to the database server
Same lateral movement pattern as in incident 1 (WinRM via wsmprovhost.exe).
Step 6: Database dump
The attacker does not create a dump, they take an existing backup from the default MSSQL backup path and compress it (T1560.001).
Six minutes later an extortion note is created, which fits the ShinyHunters pattern:
Step 7: Exfiltration
About 485 MB leave the network in one POST request.
Incident 3: AI-Enabled Malware
Kill Chain
flowchart TD
A["alex.chen installs VS Code extension via fake marketplace CDN domain"] --> B["SyntaxOptimizer-2.1.4.vsix downloaded"]
B --> C["optimizer-daemon.js (node.exe, PID 6720) starts in the background"]
C --> D["POST to generativelanguage.googleapis.com (Gemini API)"]
D --> E["AI response logged: thinking_robot_log.txt"]
E --> F["AI-generated code written: cache_worker.c"]
F --> G["Compiled to sys_worker.exe"]
G --> H["Process hollowing into svchost.exe (PID 624)"]
H --> I["credentials.db read from gcloud config (T1003.008)"]
I --> J["Exfiltration via POST to dev-metrics-cdn.com/api/v2/telemetry"]
J --> K["Lateral movement as svc_cicd_deploy to CYM-CI-RUNNER-01 (WinRM)"]
K --> L["build_production.ps1 overwritten (T1505.003 Build Hook Injection)"]
There was no classic briefing here. The entry point were GTI references about PROMPTFLUX (dropper that uses Gemini at runtime to obfuscate itself, logs to thinking_robot_log.txt) and PROMPTSTEAL (data miner that lets an LLM generate its collection commands). The environment implements the pattern in Node.js instead of VBScript.
Decoys
Several Kubernetes / GCP findings looked promising (the China-nexus section of the same GTI report describes Gemini abuse for cloud and Kubernetes work): an anonymous K8s RBAC abuse alert, a Scout Suite recon on Cloud KMS and a secret access via the known “fabric8-rbac” misconfiguration. All of them had the wrong date (13.10., 25.10. and 27.10. instead of 15.10.). The real chain was on a single developer workstation (CYM-DEV-WS-019).
Step 1: Malicious extension
The referer is the real marketplace, but the download comes from cdn-cache.net. The real marketplace hostname is only used as a subdomain prefix to look legitimate.
Step 2: The AI request
Two minutes later the extension daemon starts and calls the Gemini API:
The call comes from a programmatic client (node-fetch), not a browser, and is rated pagerisk=0 because “cloud engineer uses an AI tool” looks plausible.
Step 3: Generated code gets compiled
Right after the response, two files are written by the same node.exe process:
thinking_robot_log.txt is exactly the file name from the PROMPTFLUX “Thinking Robot” module. Later Sysmon Event 8 shows the injection:
Step 4: Cloud credential theft
13 minutes after the injection, the hollowed process (parent svchost.exe, SYSTEM) reads the gcloud credentials:
type is the Windows cat. The parent svchost.exe as SYSTEM, instead of a normal user shell, proves that this happens from the hollowed process.
Step 5: Exfiltration
Domain and path look like normal developer monitoring traffic.
Step 6: Lateral movement and persistence
With the harvested CI/CD service account a WinRM connection (port 5985) to CYM-CI-RUNNER-01 is opened. 13 minutes later on the target:
The attacker code now runs with every future build, which is very effective persistence in a DevOps environment because it looks completely legitimate. The events on the target host are stored under principal.hostname instead of target.hostname.
IOCs
| IOC | Type | Incident |
|---|---|---|
marketingtrends-insider.com |
Domain | 1, drive-by referer |
static.cdn-updater-distribution.net |
Domain | 1, payload download |
d27e15eb.exe |
File | 1, second stage |
svc_backup |
Account | 1, cracked service account |
sso-cymbal-portal.com |
Domain | 2, phishing |
157.131.172.74 |
IP | 2, attacker proxy |
918237461928-tooglebox-recall.apps.googleusercontent.com |
OAuth client ID | 2, anti-forensics |
svc_dbadmin |
Account | 2, stolen service account |
142.11.200.188:8888 |
IP:Port | 2, exfiltration |
marketplace.visualstudio.com.cdn-cache.net |
Domain | 3, fake download source |
dev-metrics-cdn.com |
Domain | 3, exfiltration |
svc_cicd_deploy |
Account | 3, stolen CI/CD account |
Lessons Learned
- A case summary (also an AI-generated one) is a summary, not evidence. Twice a concrete number (the svchost PIDs) was simply wrong, only the raw log gave the real answer.
- SOAR alert grouping bundles alerts by shared entities. A rule that is too coarse (here: the file
explorer.exe) puts unrelated incidents into the same case. - An alert closed as “Lab test, NOT_MALICIOUS” does not mean a human ever looked at it. Check the audit trail.
Execution ProcessIDin the<System>block of a Windows event is the PID of the logging service (usually Sysmon), not of the described process. The real PID is in<EventData><Data Name="ProcessId">.- UDM regex (RE2) has no trailing
/pattern/i. Case-insensitivity must be(?i)at the start of the pattern, or thenocasekeyword. - The highest risk score is not always the right lead. The real phishing hit had a lower score than two decoys because it was disguised company-specifically.
- GTI PDF reports are a reference for the TTP pattern, not for exact names and timestamps. The real in-game data differed from the examples every time.
- Google Workspace has its own log categories per
applicationName(gmail,token,user_accounts). Search for the matching UDM event type (e.g.EMAIL_TRANSACTION) instead of guessing keywords. - After process hollowing, the malicious code runs under the identity of the hollowed process. Pivot via the unique
ProcessGuid/ PID instead of guessing time windows. - If a timeline does not fit, do not fiddle with time zone conversions. Search the exact raw event by a unique ID (ProcessGuid, file ID, SID).
Possible Mitigation
- Block script file types (
.js,.vbs) from running out of the Downloads folder, and restrictwscript.exe - Enforce AES-only Kerberos and use long random passwords or gMSAs for service accounts (against Kerberoasting)
- Alert on Event 4769 with RC4 (
0x17) ticket encryption - Restrict and monitor WinRM between workstations and servers
- Phishing-resistant MFA (FIDO2), alert on new MFA factors and restrict OAuth app consent
- Never store credentials in documentation, use a secret manager
- Only allow vetted VS Code extensions and monitor non-browser clients that call AI APIs
- Protect CI/CD runners and build scripts (integrity checks, least privilege for pipeline accounts)
- Tune SOAR alert grouping so that generic entities like
explorer.exeare not used for grouping