TechPastWeek logo
Security7 min read

A Stolen-Token Campaign Pushed a CI Secret Stealer Into Thousands of GitHub Repositories

A campaign called wfx-loot used hijacked GitHub accounts to probe 61,764 repositories and plant a credential-stealing script in at least 1,835 of them.

A GitHub repository page showing the file list and the Actions tab used to run continuous integration workflows.
On this page

Key takeaways

  • Security firm Aikido said 4,837 compromised accounts opened probe pull requests against 61,764 GitHub repositories between 10 and 11 November.
  • The wfx-loot payload ran in at least 1,835 repositories, stealing CI secrets, cloud credentials and AI provider keys, Aikido said.
  • SafeDep counted 8,093 public repositories that received malicious commits, a wider measure that does not prove the code executed.
  • The attacker did not exploit a GitHub bug. It used valid stolen tokens to push workflows, npm install hooks and short-lived branches, then deleted the run logs.
  • Researchers traced 173 probe pull requests to tokens issued to the Fly.io GitHub app and said they found no evidence that Fly.io's infrastructure was breached.

A credential-theft campaign that started on 10 November used thousands of hijacked GitHub accounts to test write access to tens of thousands of repositories, then planted a secret stealer wherever the permissions allowed it. Security firm Aikido, which named the operation wfx-loot, said the payload ran in at least 1,835 repositories and exfiltrated credentials.

Related reading: AI Agents Found a Zero-Day in MONAI Label and Rooted a University Server · China-Linked Hackers Exploited a vCenter Zero-Day Weeks Before the Patch.

The scale is unusual even by the standards of software supply-chain attacks. Aikido counted 4,837 compromised accounts opening probe pull requests against 61,764 repositories. Of those requests, 10,431 went into repositories the account did not own, reached through collaborator or organization access. About 92% were still open when the firm published its analysis on 12 November.

SafeDep, a separate security firm, counted 8,093 public repositories that received malicious commits, according to SecurityWeek. That figure measures how far the code spread, not how often it ran. Aikido's 1,835 is a count of repositories where it found signs of execution. Both numbers are lower bounds.

How the attack worked

The operation did not exploit a flaw in GitHub. StepSecurity said the attacker relied on valid tokens belonging to developer and bot accounts. Aikido's timeline puts the first activity at about 12:45 UTC on 10 November, with probe pull requests continuing until 15:00 UTC on 11 November.

The probe was deliberately dull. The attacker opened a pull request from a branch called wfx-docs with the title "docs: dependency audit notes". The request changed nothing useful. Its purpose was to confirm that the stolen token could write to the repository.

Where it could, the attacker pushed one of three payloads. One added about 1,500 lines to an existing GitHub Actions workflow under the commit message "ci: refresh build cache". A second added a .ci-xxxxxx.sh script and a postinstall hook to package.json, so the script ran on any npm install. A third created a .github/workflows/build-audit-xxxxxx.yml file on a short-lived branch named chore/ plus eight random characters.

The payload was a single shell script of about 1,500 lines. Aikido described it as heavily commented and not obfuscated, with "wfx" in its log lines and variable names. It ran in four stages: collect, escalate, persist and exfiltrate.

What it took

The collection stage was broad. The script read GitHub Actions secrets and variables, the workflow's GITHUB_TOKEN, OIDC and runtime tokens, and the runner's full environment. It looked for credential files for AWS, Azure, Google Cloud, Docker, Kubernetes, npm, PyPI, Git, SSH, GPG, Vault and Terraform. It also searched for keys belonging to AI providers and developer tools, including OpenAI, Anthropic, Gemini, Hugging Face, Cursor, Copilot, Codex and Ollama.

It did not stop at files. The script read shell histories, searched recursively for .env, .pem, .p12, .jks, .kdbx and kubeconfig files, and inspected the memory of running processes, prioritizing the GitHub Actions runner processes. On self-hosted runners it also tried cloud instance metadata, Docker state, the runner's RSA parameters and systemd units.

The escalation stage turned those credentials into deeper access. The script minted GitHub OIDC tokens for AWS, Google Cloud and Azure, exchanged them for cloud sessions, and chained into other IAM roles it found in the repository. On AWS it called iam:CreateAccessKey to create long-lived credentials. On Google Cloud it signed JWTs with service account keys and impersonated service accounts. On Azure it replayed cached CLI refresh tokens against first-party Microsoft client IDs to reach Graph, ARM and Key Vault.

Persistence and cover-up

Three features were designed to survive credential rotation. Early in the run, before most of the collection, the payload posted forged success status checks against the commit, using the names of the repository's required checks. Branch protection matches on the check's context name, not on the app that posted it, so the forged checks could clear merge gates while the malware was still running.

On AWS, the script rewrote the chained role's trust policy to accept any AWS account. On Azure, it added federated identity credentials named wfx-ci-... to Microsoft Entra applications, pointing back to the compromised repository. That allowed any workflow in the repository to authenticate to Azure without a stored secret, even after every token had been rotated.

Exfiltration went to one of two HTTP endpoints, 194.150.166.244:8080/exfil or 18.143.168.96:8080/exfil, according to Aikido. The script supports other channels through GitHub's own infrastructure, including the Actions cache API, workflow step summaries and artifact uploads, but the deployed configuration in this campaign used direct HTTP posts.

Then it cleaned up. The attacker reverted the changes within at most a minute and deleted the workflow run and its logs. A missing run number in the workflow's history is often the only trace that the payload executed.

The campaign reached published software through automated release triggers. Aikido identified a malicious Packagist package, digitalelvis/neuronai-studio version 3.14.3, and two tags of the Go module github.com/onuragtas/docker-env, v1.0.278 and v1.0.280. In the Go case the payload was in a workflow file only. The Go source and release binaries were unchanged, and Go builds do not run workflow files, so consumers of the module may not have been affected.

StepSecurity traced 173 probe pull requests to tokens issued to the Fly.io GitHub app. The firm said it could not determine how the tokens were obtained and found no evidence that Fly.io's infrastructure had been breached. Fly.io has not issued a public statement in the coverage reviewed for this article.

Aikido also found overlap with GhostAction, a separate campaign in October that used secret-scanning workflows and a fake font file to steal developer credentials. Accounts hit in October were more likely to be hit again in November, and in more than 100 repositories the new payload was injected into the old GhostAction workflow files. Aikido assessed that the two operations probably involved different actors, based on differences in infrastructure, tooling and tradecraft. Indonesian-language comments in newer versions of the malware suggest a hasty edit, not the same team.

Why CI is the target

Continuous integration runners are attractive because they sit at the intersection of code, credentials and deployment rights. A workflow that builds and publishes a package usually has permission to read the repository, push artifacts to a registry and deploy to a cloud environment. In many setups it also holds a long-lived cloud key or an identity that can mint one.

The wfx-loot script was built to exploit exactly that. It did not need to break into a developer's laptop. It needed a token that could write to a repository, and then a workflow that would execute on the next push. The probe pull request told the attacker which repositories met the first condition. The existing workflow or the npm install hook supplied the second.

This is also why the campaign spread beyond the initial repositories. Automated publishing triggers took the payload into a Packagist package and two Go module tags. The attacker appears to have been after CI secrets rather than a broad software supply-chain compromise, but the package releases show how quickly a repository-level intrusion can reach downstream users.

GitHub's own controls help, but they are not a substitute for token hygiene. Branch protection, required reviews and environment secrets reduce the blast radius. They do not stop an attacker who already has a valid token and can push to a branch that runs a workflow. The forged status checks in this campaign were designed to defeat the review gate without touching the code under review.

What repository owners should do

The practical advice is blunt. If the payload ran in a repository, every secret and cloud role available to its workflows should be treated as compromised. Rotating the GitHub token is not enough if the attacker created an IAM access key or an Azure federated identity that outlives it.

Aikido's checklist includes revoking compromised tokens, rotating exposed secrets, reviewing cloud audit logs for new access keys, trust policy changes and federated identity credentials, and checking for the campaign's branch names, commit messages and pull request titles. Organizations that use self-hosted runners should also inspect the runner host for persistence.

A CI runner can reach source code, package registries, cloud accounts and, increasingly, AI provider keys. A single stolen developer token can inherit all of that. The attacker in this campaign did not need a zero-day. It needed one valid token and a repository where nobody was watching the pull request queue.

The counts are vendor estimates

The 61,764 repositories, 4,837 accounts and 1,835 executions are Aikido's figures. SafeDep's 8,093 is a count of repositories with malicious commits, not confirmed executions. No complete victim list has been published, GitHub has not confirmed the totals, and the attacker's identity is unknown. The Fly.io link is based on token issuance, not on evidence that Fly.io was breached.

Tags

githubsupply-chainci-cdcredentialsmalware

Frequently asked questions

What is the wfx-loot campaign?
It is a credential-theft operation that began on 10 November 2026. Attackers used stolen GitHub tokens to open probe pull requests in tens of thousands of repositories, then pushed a shell script that collects secrets from CI runners and cloud environments.
Was GitHub itself breached?
No evidence points to a GitHub platform vulnerability. Aikido and StepSecurity both said the attacker relied on valid tokens belonging to compromised developer and bot accounts, not on a bug in GitHub Actions.
Which repositories are affected?
Aikido identified at least 1,835 repositories where the payload ran. SafeDep counted 8,093 public repositories with malicious commits. The full list has not been published, so repository owners should check for the indicators rather than rely on a public tally.
What should affected teams do?
Assume every secret and cloud role available to the affected workflows is compromised. Rotate tokens and credentials, revoke sessions, review cloud audit logs for new access keys or federated identities, and check for the branch names, commit messages and pull request titles listed by the researchers.
Share

Found this useful? Share it with your team.

TPW Editorial

TPW Editorial

Editorial Team

TechPastWeek's editorial team reports and edits every story we publish. We research each piece against primary and credible secondary sources, and a human editor reviews and signs off before it goes live.

Keep reading