OpenSourceMalware has identified a new stage in the ongoing "GhostAction" malicious campaign that GitGuardian originally identified in 2025. This latest evolution appears to be an escalation and used two new domains targeting GitHub and GitLab. The GhostAction campaign is a continuing supply-chain campaign in which attackers compromise GitHub accounts and inject malicious Actions workflows across every repository those accounts can modify. Triggered by pushes or manual execution, the workflows steal CI/CD secrets—including package-publishing, cloud, GitHub, and AI-service credentials—and exfiltrate them to attacker-controlled infrastructure. Later variants expanded collection by scanning repository contents and complete git history, recovering credentials that developers believed had been deleted. GitGuardian ( returns/)
Earlier today Socket and StepSecurity subsequently traced a renewed, highly automated wave to compromised maintainers whose access exposed hundreds of repositories, including prominent projects and long-dormant codebases. Their findings show that GhostAction remains an active account-takeover and credential-theft operation: attackers enumerate all writable repositories, push workflows directly without review, trigger executions, and use stolen publishing or cloud secrets to enable further supply-chain compromise.
The established GhostAction payload reads specifically named GitHub Actions secrets and sends them to attacker infrastructure. The newer .github/workflows/security-audit.yml variant also searches the complete checkout and git history for credential-shaped strings, captures contextual lines surrounding AWS keys, and POSTs the combined results over unencrypted HTTP to 193.32.204[.]199/?c=monami .
Responders must assume exposure of active Actions secrets and credentials previously committed and later deleted. Removing the workflow does not revoke stolen credentials, invalidate the GitHub token used to alter the repository, or address packages, images, releases, and deployments produced during the compromise.
Early Warning: A Possible New Developer-Targeting Wave
Two newly registered, code-hosting-themed domains may signal the phase of developer targeting. my-gitlab[.]com was registered on September 22, 2026, followed 17 days later by my-github[.]com . The domains the exact my- .com construction and the parallel .my- .com form. my-github[.]com points directly to the active GhostAction collector IP; gitlab.my-gitlab[.]com hosts an apparent GitLab service on separate AWS infrastructure. Common ownership is not yet proven, but the naming, timing, and developer-facing services justify early monitoring for phishing, malicious clone URLs, token theft, CI configuration abuse, and payload delivery.
The infrastructure suggests a possible shift from conspicuous IP-address collectors toward domains designed to look familiar to developers:
This pattern could support several attacks against developers: GitHub or GitLab credential phishing, OAuth or personal-access-token theft, malicious repository clone instructions, fake API endpoints, poisoned package or release downloads, runner registration, and CI/CD secret collection. These are forecast scenarios derived from the naming and exposed services; they have not all been observed.
my-github[.]com is the higher-confidence indicator because it resolves directly to 193.32.204[.]199 , the current GhostAction exfiltration and scanning host. It also has wildcard DNS, allowing whoever controls the domain to use convincing hostnames without publishing individual records. my-gitlab[.]com is a lower-confidence watchlist domain: its configured gitlab.my-gitlab[.]com hostname resolves to an AWS EC2 address in Hong Kong where passive Shodan data identifies GitLab and nginx on ports 80 and 443.
The pair is not technically attributed to one operator. They use different registrars, registrant records, Cloudflare nameserver pairs, IP addresses, ASNs, and DNS designs, and no shared certificate or account artifact has been recovered. Defenders should nevertheless hunt and monitor both domains now because waiting for confirmed victim telemetry would forfeit the value of the early warning.
Recommended monitoring:
DNS, proxy, browser, email, and endpoint events containing either apex domain or any subdomain;
DNS, proxy, browser, email, and endpoint events containing either apex domain or any subdomain;
Git remotes, package metadata, documentation, workflow files, and shell history referencing either domain;
Git remotes, package metadata, documentation, workflow files, and shell history referencing either domain;
authentication pages, OAuth redirects, personal-access-token prompts, runner-registration instructions, and clone URLs using the domains;
authentication pages, OAuth redirects, personal-access-token prompts, runner-registration instructions, and clone URLs using the domains;
outbound connections from developer workstations, CI runners, build systems, and package-publishing hosts;
outbound connections from developer workstations, CI runners, build systems, and package-publishing hosts;
future DNS, certificate, hosting, and repository changes that create a direct link between the two domains.
future DNS, certificate, hosting, and repository changes that create a direct link between the two domains.
CI/CD credential theft and software supply-chain intrusion
GitHub and GitHub Actions
Critical where workflows can access publishing, cloud, or deployment secrets
193.32.204[.]199 over HTTP
github_actions_security.yml , security-check.yml , security-audit.yml
push , workflow_dispatch
OpenSourceMalware identified a malicious PyPI package claudecord and reached out to the maintainer directly. We worked with that maintainer to help them mitigate the malware that had compromised dozens of their repositories. During that engagement, we noticed that there were hundreds of GhostActions victims with a new GitHub Actions based payload. Version 0.3.2 of the claudecord PyPI package contains .github/workflows/github_actions_security.yml , which POSTs named deployment, npm, and PyPI secrets to . Its SHA-256 is:
The claimed upstream, kanavdhanda/claudeCord , shows the initial implant in commit a69f8c4 , followed two seconds later by Trigger security scan . Commit dd08e03 added .github/workflows/security-audit.yml the day. Commit 7e03c5a later removed the remaining malicious workflow.
GitGuardian reported 772 affected repositories between August 31 and September 30. OpenSourceMalware can currently recover 717 through that cutoff and 790 through October 9. Repositories and commits may disappear after disclosure, while injections continued after GitGuardian’s window; the present dataset is therefore a reproducible public lower bound.
October 8 Mass Injection
Socket and StepSecurity resolved the newest activity to two compressed sweeps:
kitao/pyxel received one add and two update commits
39 source repositories, 279 forks, plus uber/athenadriver
Automated enumeration of writable repositories
The sweep included active projects and repositories dormant for roughly a decade, supporting automated enumeration rather than project selection. kitao/pyxel had approximately 18,420 stars at Socket’s collection time. The Uber-owned uber/athenadriver repository was reachable because its original author retained write access after the project moved under the Uber organization.
The athenadriver injection went directly to master without a pull request or review and used the legitimate maintainer identity as author and committer. StepSecurity reports that the commit was unsigned. Author identity therefore provides weak detection when a valid credential is abused; content, review state, burst timing, push path, and runner egress are stronger signals.
Socket observed successful executions and found the workflow still present on default branches it checked on October 9. It had not observed malicious PyPI or crates.io releases attributable to this wave at publication time. That narrows observed impact but does not reduce the need to rotate publishing credentials.
Reported repositories
bold-dhawan.45-139-104-115.plesk[.]page
~900 total for the wave
objective-hopper.45-139-104-115.plesk[.]page
Oct–Dec 2025; March 2026
Nov 2025; March–April 2026
193.32.204[.]199:3000/api/workflow/receive?inj=
193.32.204[.]199/?c=monami
StepSecurity places the current IP in honeypot telemetry on September 4, 2026 and in an infected public repository on September 5. Those dates provide an earlier investigation boundary than the major public injection bursts.
Infrastructure Reuse Beyond GhostAction
At OpenSourceMalware our years of incident response in software supply chain attack led us to look deeper at the indicators and behaviours of the GhostAction threat actor. In particular we like to go deep on the infrastructure used as it often is the best way to pivot into new, previously unknown, parts of the campaign. In this case, dedicated infrastructure research found that 193.32.204[.]199 is also being used as active malicious internet scanner. GreyNoise, Shodan, OTX, SCARD, SANS ISC, and BlockList.de independently observed scanning, brute-force, web probing, or honeypot traffic. A SANS daily view recorded 11,286 probes to TCP/8080, while SCARD recorded 413 events dominated by suspicious web-scanner user agents. The threat actor, or actors, is clearly using this Ubuntu server to act both as a inbound exfiltration collection server, AND as a outbound scanner looking for vulnerabilities and exploitable misconfigurations.
It is very unusual to see infrastructure used in software supply chain attacks re-used in this way.
In a similar way, the legacy IP used in the first version of GhostAction identified by GitGuardian, 45.139.104[.]115 has 389 OTX passive-DNS records spanning a phishing-heavy neighborhood. Recurring themes include Ameli/Carte Vitale, government fines, parcel delivery, Netflix, banking, and SNCF. carte-avantage[.]com was independently reported as an SNCF payment-card phishing site before its 2025 GhostAction use. This establishes multi-purpose malicious use of the infrastructure but does not prove that one operator controlled every co-hosted domain.
Current IP ownership and exposure
RIPE RDAP assigns 193.32.204.0/24 to the object vcyber , with Vigilant Cyber SAS as the registered organization and abuse@vigilantcyber.top as the abuse . The prefix is currently announced by AS153622, Madina IT. Commercial geolocation sources variously place the address in Helsinki, Istanbul, or Turkey. These records describe allocation, routing, and database-derived location respectively; none establishes where the GhostAction operator resides.
Shodan InternetDB observed OpenSSH 8.9p1 on TCP/22, Apache 2.4.52 on TCP/80, and TCP/8900 as active on the IP address. URLScan independently recorded returning Apache 2.4.52 on Ubuntu. Version-derived CVEs in Shodan are exposure hypotheses and do not prove exploitability.
my-github[.]com : same-day domain on the current collector
my-github[.]com is a new and highly relevant pivot on the active GhostAction IP:
Dynadot Inc., IANA ID 472
Not disclosed in public RDAP
kianchau.ns.cloudflare.com , selah.ns.cloudflare.com
Not signed in observed registration data
The domain was registered on the day this infrastructure was being publicly investigated and uses a name that impersonates or invokes GitHub. It does not use Cloudflare’s reverse proxy in the observed A record; DNS resolves directly to the GhostAction collector. Public scan reporting associated the domain with Vigilant Cyber hosting and observed a valid TLS presentation.
URLScan submitted at 15:48 UTC on October 9. The observed finished at a YouTube Rickroll URL. That behavior may represent operator taunting, a disposable redirect, an unrelated tenant, or researcher activity after disclosure. It is not evidence of credential phishing by itself.
The timing, brand-related name, and exact A-record overlap make my-github[.]com a high-priority indicator and pivot. There is still no direct workflow, account, registrar, or server-side evidence proving that the GhostAction operator registered it. The report therefore classifies it as suspicious infrastructure associated by IP, operator attribution unresolved .
Any outbound GitHub runner connection to my-github[.]com should be investigated as potential campaign adaptation even if the workflow no longer contains the literal IP.
Four reported labels— ghassets.my-github[.]com , github.my-github[.]com , api.my-github[.]com , and gh.my-github[.]com —also resolve to 193.32.204.199 . None currently has an AAAA, CNAME, or TXT record, and no exact public URLScan capture or OTX passive-DNS history was found for them.
Random control labels also resolve to the same address with the same TTL, confirming wildcard DNS for *.my-github.com . The names are semantically consistent with GitHub impersonation, but DNS resolution does not prove they are separately configured services: any invented label resolves. Count the apex as the infrastructure node, retain the four surfaced labels as hunt terms, and require HTTP Host , TLS SNI, certificate, email, workflow, or victim telemetry before asserting active use of an individual subdomain.
my-gitlab[.]com was registered on September 22, 2026, 17 days before my-github[.]com . It follows the same my- .com naming pattern, making it a useful pivot, but current infrastructure does not directly connect it to GhostAction. It uses Gname rather than Dynadot, has a different Cloudflare nameserver pair, and does not resolve to 193.32.204[.]199 .
The apex currently has no address or mail records. The explicitly configured gitlab.my-gitlab[.]com hostname resolves to 18.167.164[.]26 , an AWS EC2 address in Hong Kong. Shodan InternetDB reports ports 80 and 443 with GitLab and nginx fingerprints, consistent with a functioning self-hosted GitLab service. Random subdomain controls return NXDOMAIN, so this domain does not use the wildcard behavior seen on my-github[.]com . Public URLScan, OTX, , and Certificate Transparency checks produced no malicious or campaign-specific evidence.
We want to stress that my-gitlab[.]com and gitlab.my-gitlab[.]com are not yet proven to be malicious . Their naming and timing warrant monitoring, but treating them as confirmed GhostAction infrastructure would exceed the evidence. If OSM identifies that my-gitlab[.]com is benign, we will remove it from the indicator list.
The strongest connection between my-gitlab[.]com and my-github[.]com is the exact my- .com naming construction combined with registration 17 days apart. There is also a parallel .my- .com form: gitlab.my-gitlab[.]com is explicitly configured, while github.my-github[.]com resolves through the GitHub-themed domain’s wildcard. Tests of analogous GitHub, GitLab, API, asset, registry, package, authentication, and login labels under my-gitlab[.]com returned NXDOMAIN except for gitlab.my-gitlab[.]com ; the evidence therefore shows a strong lexical pattern, not a mirrored subdomain deployment.
No operator-specific link was found: the domains use different registrars, privacy/registrant records, Cloudflare nameserver pairs, SOA serials, IP addresses, ASNs, DNS designs, and hosting providers. No shared certificate, passive observation, or relevant indexed co-occurrence was identified. Cloudflare DNS, one-year registration, transfer locks, and disabled DNSSEC are common defaults and carry little attribution weight. The relationship remains a credible hypothesis until a shared token, certificate or key, origin, account, repository reference, payload, or victim-side sequence is recovered.
carte-avantage[.]com : phishing-to-GhostAction crossover
Public victim reports describe sncf.carte-avantage[.]com impersonating SNCF Connect and offering a €49 discount card for €2.45. Reports state that the site collected identity and payment-card data while most non-payment was inert. GitGuardian later found carte-avantage[.]com used as a GhostAction exfiltration endpoint.
Current RDAP shows a Dynadot registration from August 15, 2025, an undisclosed registrant, and redemption status after expiration in August 2026. The domain currently returns NXDOMAIN. It received Let’s Encrypt certificates immediately after the 2025 registration, resolved to 45.139.104[.]115 by August 17, and appeared in GhostAction the following month. That sequence strongly ties the 2025 registration instance to the legacy campaign infrastructure.
Archives and Certificate Transparency show older domain lives in 2018, 2021–2022, and 2024. The 2025 creation date reflects re-registration rather than first-ever use. Consequently, the 2022/2024 SNCF phishing operator and 2025 GhostAction operator cannot be assumed identical without historical WHOIS, registrar-payment, or server-control evidence. Public RDAP does not expose a defensible individual owner.
Full ownership, routing, passive-DNS, certificate, scanning, and attribution analysis is in ghostaction_infrastructure_2026-10-09.md .
Valid repository access: the actor uses access associated with the victim identity. Displayed authorship alone cannot distinguish authorized work from token abuse.
Valid repository access: the actor uses access associated with the victim identity. Displayed authorship alone cannot distinguish authorized work from token abuse.
Workflow injection: a security-themed file is written below .github/workflows/ .
Workflow injection: a security-themed file is written below .github/workflows/ .
Execution: the injection push can trigger the newly added workflow; later pushes retrigger it, and workflow_dispatch permits manual execution.
Execution: the injection push can trigger the newly added workflow; later pushes retrigger it, and workflow_dispatch permits manual execution.
Retriggering: an inert timestamp is appended to README.md , creating another push.
Retriggering: an inert timestamp is appended to README.md , creating another push.
Collection: named Actions secrets are expanded; the new variant also scans current and historical source.
Collection: named Actions secrets are expanded; the new variant also scans current and historical source.
Exfiltration: curl --data-binary sends a victim-labelled multiline record over HTTP.
Exfiltration: curl --data-binary sends a victim-labelled multiline record over HTTP.
GitHub resolves ${{ secrets.NAME }} before Bash receives the script. Missing secrets become empty strings; existing values become command data. Log masking does not prevent the runner from transmitting the underlying value.
Targeted-secret workflow
The PyPI artifact contains the established collector:
Secret names vary by victim, indicating that the actor inspected existing workflow or configuration references and generated a tailored collector. It cannot enumerate stored secret values; it references names learned beforehand and relies on Actions expression expansion.
New security-audit.yml
The latest version combines three collectors:
explicit repository-specific Actions secrets;
explicit repository-specific Actions secrets;
regex matches in the checked-out tree;
regex matches in the checked-out tree;
regex matches and contextual lines from all reachable git patches.
regex matches and contextual lines from all reachable git patches.
It sends results to . The query value may label a campaign or collector version; its server-side meaning is unresolved.
Full-history checkout
fetch-depth: 0 obtains complete reachable history instead of a shallow checkout. Credentials deleted from the visible branch remain available to git log -p --all . The setting is legitimate by itself and becomes high-risk when correlated with secret regexes and an unrelated external POST.
Victim identity and direct secrets
GITHUB_REPOSITORY identifies the victim. Because literal parameter names remain even when all secrets are empty, the -n condition is always true. The example targets Rust and Python publishing plus a general access token; observed lists also include cloud, container, SSH, database, bot, npm, PyPI, and GitHub credentials.
Current-tree collection
-r scans recursively, -E enables extended regexes, -i ignores case, -h removes filenames, and -o returns only matches. Errors are hidden and results deduplicated. Targeted families include:
AWS long-term and STS access-key IDs
ghp_... , github_pat_...
AWS secret access key assignment
AWS session-token assignment
The regex uses PCRE non-capturing groups such as (?:v1-)? with grep -E , which implements POSIX ERE. Behavior may vary and produce warnings or missed matches; 2>/dev/null hides failures. This defect reduces reliability but does not neutralize ordinary ERE branches.
Git-history collection
git log -p --all emits metadata and diffs for every reachable reference. The actor retains up to 200,000 output lines and 300 unique matches. Deleted lines are included, so credentials removed from current source can still be collected. Large repositories bias toward recent history because git normally walks newest commits first.
AWS context collection
These commands collect two surrounding lines rather than only the key ID. Context may disclose the paired secret key, region, role, account, bucket, hostname, username, or other credentials. Current and historical context are each capped at 150 lines and wrapped with distinctive AKIA_CTX_START and AKIA_CTX_END markers.
--data-binary preserves newlines. The body contains repository identity, named secrets, AWS context, current matches, and historical matches. Plain HTTP exposes stolen data in transit. -s suppresses output, -m 20 limits delay, and || true prevents a failed POST from failing the step. The final non-empty check is always satisfied because the body begins with the repository name, so every run sends at least a victim beacon.
GitHub Hunting Queries
High-confidence infrastructure searches
The most effective primary quotes the IP and scopes the result set to executable GitHub Actions workflow paths, independent of filename:
This query detects github_actions_security.yml , security-check.yml , security-audit.yml , and renamed copies in one pass. Quoting the IP requires the complete literal indicator, while the path constraint removes documentation, IOC feeds, issue templates, and non-executable source files that mention the campaign. Results still require content review because defenders may intentionally commit detections or blocked indicators inside workflow files.
Use the following searches as broader discovery and filename-specific pivots:
Version-specific payload searches
Historical infrastructure
Phrase also returns longer messages and bodies. Fetch the file at each returned SHA and verify infrastructure or a distinctive content marker before assigning it to the campaign.
Known IP plus malicious workflow path
Direct campaign correlation
AKIA_CTX_START plus git log -p --all
Distinctive latest collector
?c=monami plus --data-binary
Version-specific route
Exact commit message alone
fetch-depth: 0 plus secret regexes
Legitimate scanners do this
security-audit.yml filename alone
Common legitimate filename
October 8 Account and Repository Pivots
Treat fork results separately from source repositories. Socket counted 279 affected forks under henrywoo ; a committed workflow in a fork becomes an execution risk when Actions is enabled and a qualifying event gives it access to that fork’s secret context.
Validated Repository Scope
Exact message; historical file and IP verified
Updated main workflow
Exact message; historical file and IP verified
Historical file and endpoint verified
Union of validated evidence
Supporting datasets are ghostaction_repositories_summary.csv , ghostaction_affected_repositories.csv , and ghostaction_trigger_commits_sample.csv beside this report.
Valid Accounts: Cloud Accounts
Compromise Software Supply Chain
Event Triggered Execution
File and Directory Discovery
Data from Local System
Exfiltration Over C2 Channel
Public evidence does not establish how the original GitHub credentials were obtained.
Indicators of Compromise
Filenames alone are not malicious indicators; correlate them with infrastructure or behavior.
The IP is shared cloud infrastructure and should only be used with the hostname or other corroborating context.
Separate DevOpsGPT Miner
Commit 614a565 modifies Dockerfile , run.sh , and two scheduler files to install and persist XMRig as /usr/local/bin/pyworker . Static XOR decoding with devops2024 yields:
The five expected fields are present, differing from GitGuardian’s conclusion that the configuration was truncated. The miner predates GhostAction in this repository and uses different, tailored tradecraft. Treat it as separate activity through a shared compromised identity unless further evidence connects the operators.
Disable Actions temporarily and preserve workflows, run IDs, logs, commits, audit logs, and package/release records.
Disable Actions temporarily and preserve workflows, run IDs, logs, commits, audit logs, and package/release records.
Revoke GitHub PATs, OAuth grants, App credentials, deploy keys, and sessions associated with the compromised identity.
Revoke GitHub PATs, OAuth grants, App credentials, deploy keys, and sessions associated with the compromised identity.
Remove malicious workflows from executable branches and tags; preserve evidence separately.
Remove malicious workflows from executable branches and tags; preserve evidence separately.
Block campaign infrastructure without contacting it.
Block campaign infrastructure without contacting it.
Rotate every named Actions secret, then and rotate matching credentials in full git history.
Rotate every named Actions secret, then and rotate matching credentials in full git history.
Prioritize source-control, registry, cloud, SSH, deployment, database, and container credentials.
Prioritize source-control, registry, cloud, SSH, deployment, database, and container credentials.
Audit packages, releases, images, and deployments produced during the compromise window.
Audit packages, releases, images, and deployments produced during the compromise window.
Rebuild trusted artifacts from a verified pre-compromise commit in a clean environment.
Rebuild trusted artifacts from a verified pre-compromise commit in a clean environment.
Harden repositories with protected branches, reviewed workflow changes, .github/workflows/** ownership rules, least-privilege organization secrets, short-lived credentials, and cloud OIDC. Alert when workflows combine full-history checkout, git-diff scanning, credential regexes, and outbound network clients.
Require approval for new or modified workflows wherever repository policy permits it. StepSecurity reports that workflow approval was the control observed stopping exfiltration in this wave. Apply runner egress allowlists as a second layer: the collector connects directly to a raw IP on ports 80 or 3000, producing no DNS lookup for domain-only controls to inspect. Historical network telemetry for those destinations can identify the repository, workflow, job, and step that attempted collection.
Deleted, private, force-pushed, and unindexed repositories are outside the recoverable public set.
Deleted, private, force-pushed, and unindexed repositories are outside the recoverable public set.
The prevalence of the newest exact security-audit.yml variant remains unresolved because its generic commit message creates substantial noise.
The prevalence of the newest exact security-audit.yml variant remains unresolved because its generic commit message creates substantial noise.
Workflow content proves attempted collection, not successful delivery or credential validity.
Workflow content proves attempted collection, not successful delivery or credential validity.
c=monami and inj= appear to route or label collection; attacker infrastructure was not contacted.
c=monami and inj= appear to route or label collection; attacker infrastructure was not contacted.
Additional malware may exist under unrelated names or commit messages.
Additional malware may exist under unrelated names or commit messages.
GitGuardian GhostAction report
GitGuardian GhostAction report
StepSecurity: GhostAction Returns
StepSecurity: GhostAction Returns
Socket: New GhostAction Wave Hits Hundreds of Repositories
Socket: New GhostAction Wave Hits Hundreds of Repositories
GreyNoise record for 193.32.204.199
GreyNoise record for 193.32.204.199
SANS ISC TCP/8080 scanner telemetry
SANS ISC TCP/8080 scanner telemetry
SCARD suspicious-source telemetry
SCARD suspicious-source telemetry
SNCF phishing reports for carte-avantage.com
SNCF phishing reports for carte-avantage.com
Verisign RDAP for carte-avantage.com
Verisign RDAP for carte-avantage.com
GitHub: Using secrets in Actions
GitHub: Using secrets in Actions
MITRE ATT&CK T1195.002
MITRE ATT&CK T1195.002
MITRE ATT&CK T1552.001
MITRE ATT&CK T1552.001
GitHub : IOC in YAML
GitHub : IOC in YAML
GitHub : quoted IOC under .github/workflows
GitHub : quoted IOC under .github/workflows
GitHub : main workflow
GitHub : main workflow
GitHub : new audit workflow
GitHub : new audit workflow
OpenSourceMalware GhostAction infrastructure assessment
OpenSourceMalware GhostAction infrastructure assessment
Report suspicious packages and repositories to OpenSourceMalware.com .
Report generated by the OpenSourceMalware Team.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
