Understand how the wfx-loot campaign stole CI secrets and cloud credentials via GitHub Actions. The article brings together indicators and investigation checklists for GitHub, AWS, Azure, GCP, Vault, and Kubernetes, explains persistence mechanisms, and presents adjustments to limit the scope of pipelines.

SRE/Kubernetes Admin
Mateus Caruccio

Between November 10 at 12:45 UTC and November 11 at 15:00 UTC, 4,837 compromised GitHub accounts opened 61,839 probing pull requests across 61,764 repositories. Wherever the account had write access, a secret-stealing script ran inside GitHub Actions. Aikido Security discovered the campaign and named it wfx-loot. According to the company, the script ran in at least 1,806 repositories and sent CI secrets and cloud credentials to the attacker.
The counts vary by source. Another Aikido tally reports 1,835 executions, and the company itself says the number may be underestimated. SafeDep counted 8,093 public repositories containing malicious commits, a figure that does not indicate execution. Of the probes, 10,431 targeted repositories the account did not own but could access through collaborator or organization permissions.
No one has yet confirmed how the credentials were obtained. One victim reported a leaked GitHub CLI OAuth token. StepSecurity says the attacker uses valid tokens and does not exploit a GitHub Actions vulnerability. Around one-fifth of the accounts had already been compromised in October’s GhostAction campaign, according to GitGuardian.
How the attack works
The script does not need a vulnerability: it runs because the person who made the commit had permission to do so. The sequence described by researchers has five steps.
Probing. The compromised account creates the
wfx-docsbranch and opens a pull request titled “docs: dependency audit notes,” adding the harmless filedocs/wfx-audit.md. If the PR goes through, the attacker knows they have write access.Injection. The roughly 1,500-line script is introduced in three ways: inside an existing workflow (7,309 repositories), through an npm installation hook, or through a new workflow on a temporary branch.
Collection. It collects Actions secrets and variables,
GITHUB_TOKEN, OIDC and runtime tokens, and the runner process’s memory. It also searches for AWS, Azure, and GCP credentials, Docker and Kubernetes configuration files, SSH keys, AI API keys, and shell history.Escalation. It requests GitHub OIDC tokens with AWS, Azure, and GCP audiences and exchanges them for cloud sessions wherever the repository is trusted. It then chains role assumptions and reuses tokens against HashiCorp Vault and any Kubernetes API servers it finds in kubeconfig files.
Cleanup. The attacker typically reverts the change within minutes using “ci: restore” or “ci: cleanup” commits and deletes the workflow run within about a minute. Data is exfiltrated through HTTP POST and Telegram, directly to the attacker’s servers.
Step 5 changes the investigation. Because the run is deleted and nothing remains in caches or artifacts, the absence of traces proves little. Aikido considers an attacker commit without a corresponding workflow run number a more reliable indicator.
Why rotating secrets is not enough
The standard response to a secret leak is to rotate everything. Against wfx-loot, that leaves doors open because the payload creates new access paths that do not depend on the stolen secrets.
Platform | What the payload leaves behind |
|---|---|
AWS | Long-lived access keys created with |
Azure / Entra ID | Federated identity credentials named |
GCP | No dedicated persistence mechanism reported; collected service account keys are used to sign JWTs and impersonate accounts |
GitHub | Forged successful status checks on protected branches |
The Azure federated credential is the clearest example. With it, any workflow in the compromised repository can authenticate to Azure again without a stored secret, even after all tokens have been rotated. Aikido also considers any account that merely opened the probing PR compromised, even if the script never ran.
GitHub threat-hunting checklist
The scope includes every public and private repository and organization where anyone on the team has write access, starting November 10. The script includes configuration for private repositories, so public searches are not enough.
Search for the probing PR:
gh search prs "docs: dependency audit notes" --owner <org>.Look for branches named
wfx-docs,chore/followed by eight characters, andci-check-followed by four digits: rungit ls-remote --heads originand filter the results.Search for the commits “ci: refresh build cache,” “chore: add ci helper,” “chore: add dependency audit workflow,” “ci: restore,” and “ci: cleanup”:
git log --all --since=2026-11-10 --grep="<message>".Search workflows, scripts, and
package.jsonfiles for the script’s SHA-256 hash,89695abb392d5cfd4972672961e644bb46baa98aa1ed7173c2a2816c762735cf, and the campaign token,b42d70a7aab9140ca19e8249.Review
preinstallandpostinstallhooks added topackage.jsonsince November 10.Cross-reference commits since November 10 with the run history (
gh run list --json headSha,createdAt). A commit that should have triggered a workflow but has no corresponding run is the strongest sign that the script ran and deleted its traces.Check for workflows that pass the entire secrets object through
toJSON(secrets), the pattern used by the payload.For each suspicious commit author, revoke tokens, OAuth authorizations, SSH keys, and sessions, and require 2FA.
Close the probing PRs, delete the branches, and confirm that the revert commit actually removed the payload.
If the payload ran, treat every secret and cloud role accessible to that repository’s workflows as compromised, and continue with the cloud sections below.
AWS threat-hunting checklist
In AWS, the payload exchanges the GitHub OIDC token for a role session, creates a long-lived access key, and rewrites a role so that any AWS account can assume it. Start with roles that trust GitHub’s OIDC provider (token.actions.githubusercontent.com).
In CloudTrail, look for
AssumeRoleWithWebIdentityorAssumeRoleevents with awfx-*session name, followed byCreateAccessKeyorUpdateAssumeRolePolicy. In CloudTrail Lake or Athena, filter foruserIdentity.arncontaining/wfx-.List IAM keys created since November 10:
aws cloudtrail lookup-events --lookup-attributes AttributeKey=EventName,AttributeValue=CreateAccessKey --start-time 2026-11-10T00:00:00Z. Remove any key whose owner cannot be identified.Repeat with
AttributeValue=UpdateAssumeRolePolicyand review every modified role.Export the trust policies (
aws iam get-account-authorization-details --filter Role) and search for"Principal": {"AWS": "*"}. Restore the previous policy from the history or infrastructure code.For roles federated with GitHub, verify that the
token.actions.githubusercontent.com:subcondition restricts the repository and branch or environment. A wildcard in the repository restriction allows any workflow to assume the role.Starting from each
wfx-*session, follow the chain: which other roles were assumed, and what was read from Secrets Manager, SSM Parameter Store, and S3.Rotate secrets only after removing keys and restoring trust policies. Reversing that order lets the attacker return through the access path left open.
Azure / Entra ID threat-hunting checklist
In Azure, the payload exchanges an OIDC token or client secret for management tokens and reuses FOCI refresh tokens from cached Azure CLI sessions to access Graph, ARM, and Key Vault. Persistence comes from a wfx-ci-<n> federated credential that trusts the compromised repository.
List federated credentials across all applications and look for
wfx-ci-:for app in $(az ad app list --all --query "[].appId" -o tsv); do az ad app federated-credential list --id $app --query "[?starts_with(name,'wfx-ci-')]"; done.In Entra ID audit logs, look for application changes that added a federated identity credential since November 10, regardless of its name. The prefix is what the sources reported; an attacker can change it.
For each legitimate federated credential, verify the subject: the repository, branch, or environment should match what you expect.
In sign-in logs, look for use of the Azure CLI client ID from IP addresses or devices that are neither runners nor team machines.
Review Key Vault access since November 10 by CI-related identities.
Remove
wfx-ci-*credentials before rotating secrets, for the same reason as in AWS.
GCP threat-hunting checklist
For GCP, the sources describe escalation rather than a dedicated persistence mechanism. The payload signs JWTs using service account keys it finds and impersonates accounts through generateAccessToken and signJwt. Some analyses recommend searching for new service account keys, which is worth doing even without a confirmed report of key creation.
Confirm that Data Access audit logs for the IAM Credentials API are enabled. Without them,
GenerateAccessTokenandSignJwtcalls do not appear.Look for these calls from principals in GitHub’s workload identity pool:
gcloud logging read 'protoPayload.methodName=("GenerateAccessToken" OR "SignJwt") AND protoPayload.authenticationInfo.principalSubject:"workloadIdentityPools"' --freshness=8d. A GitHub OIDC principal you did not configure is a direct indicator.Search for service account key creation since November 10:
protoPayload.methodName="google.iam.admin.v1.CreateServiceAccountKey". Delete any keys no one can account for.Review new grants of
roles/iam.serviceAccountTokenCreatorandroles/iam.workloadIdentityUser.Review the attribute conditions on the pool’s GitHub provider (
attribute.repository,attribute.ref). Without a condition, any GitHub repository can exchange a token.Rotate service account keys that were stored in CI secrets. If the organization allows it, migrate these workflows to workload identity and block key creation.
Vault, Kubernetes, and published packages
The payload reuses OIDC tokens against HashiCorp Vault and API servers found in kubeconfig files. For Kubernetes operators, a CI runner with a deployment kubeconfig is the shortest path to the cluster.
In Vault, search for JWT authentication logins using tokens from GitHub’s OIDC issuer since November 10, and list the secrets read during those sessions.
Identify which workflows have a kubeconfig, a service account token, or OIDC access to a cluster. Those clusters are in scope.
In API server audit logs, look for requests from these identities outside normal deployment times and patterns:
ClusterRoleBindingcreation, bulk reads of Secrets, privileged pods, or pods usinghostPath.Rotate cluster credentials that were stored in CI secrets and restrict the deployment user’s RBAC permissions to the namespace it needs.
The payload also reached packages published through automation: the Packagist package neuronai-studio 3.14.3 and the Go modules docker-env and gocurl. For Go, the payload was confined to a workflow file, which Go builds never execute. Aikido recommends treating any machine that installed or opened the Packagist package as compromised.
Search the SBOMs for production images and builds for
neuronai-studio,docker-env, andgocurl. With an SBOM generated at build time, this is a search. Without one, each image must be scanned individually.Anyone who ran
npm installon an affected commit should treat the machine as compromised and rotate local keys.
What to change in your pipelines afterward
The attacker got in with valid credentials, so the damage is limited by what each workflow can reach when it runs.
Restrict OIDC trust by repository and branch or environment across all three clouds. The production deployment role should accept tokens only from the main branch or an environment that requires approval.
Remove permissions to create IAM keys, modify trust policies, and add federated credentials from CI roles. No build pipeline needs these permissions.
Declare minimal
permissions:in each workflow and expose only the secrets each job uses, rather than passingtoJSON(secrets).Require reviews for changes to
.github/workflowsand installation hooks, including changes from internal collaborators.Restrict runners’ outbound network access. wfx-loot sent data directly to the attacker’s servers over HTTP and through Telegram.
Monitor machine identities for behavioral changes: a key previously used only by a runner appearing from another country, or a workload that never accessed a vault suddenly reading from it.
Generate an SBOM at build time and store it alongside the image, so you can answer “Am I using the affected package?” when the next campaign’s package list is released.
Sources
Help Net Security: Check your GitHub repos (November 12)
Expert Insights: Hijacked GitHub Accounts Push CI Secret Stealer (November 12)
Cloud Security Alliance: The wfx Campaign (November 12)
Endor Labs: Hijacked GitHub Accounts Push wfx
Chainguard: wfx-loot
CSO Online: Thousands of GitHub repos hit (November 13)
SecurityWeek: Attacker targets 60,000 GitHub repositories
Rescana: indicators and queries
darkdive: wfx-loot
SailPoint: machine identity detection
dev.to: summary with Aikido’s recommendations
TechRadar (November 13)
CySecurity News (November 13)
pasqualepillitteri.it
Elastic: GCP Service Account Key Creation
Newsletter Getup.
Atualizações sobre Kubernetes e Software Supply Chain Security todos os meses.
Operating Kubernetes in production for more than 13 years. With Quor, this experience extends to software supply chain security as well.
GET UP
© Getup · 2026

