Skip to main content
All incident reports
Incident report

Stolen GitHub accounts used to plant secret stealer in build workflows

An attacker used stolen credentials for 4,837 GitHub accounts to plant a secret stealer in repositories' continuous integration (CI) workflows, Aikido Security says. Aikido says the stealer ran in at least 1,806 repositories, while SafeDep counted malicious commits in 8,093 public repositories.

· First published

The short version

  • The campaign. Aikido Security says an operation it calls wfx-loot began on 10 November using stolen credentials for 4,837 GitHub accounts. It opened probe pull requests in 61,764 repositories and the stealer ran in at least 1,806, Aikido says.
  • The scale. SafeDep counted 8,093 public repositories that received malicious commits between 10 and 12 November. It searched only default branches of public repositories, so it says the real total is higher.
  • The method. The attacker used pull requests titled docs: dependency audit notes from wfx-docs branches to test write access, then added the stealer to workflows or npm install hooks. It usually reverted the change and deleted workflow runs and logs.
  • The target. The stealer was built to collect repository secrets, tokens, cloud credentials, AI service keys and runner memory, and can exchange OpenID Connect (OIDC) tokens for AWS, Azure and Google Cloud sessions. It also has mechanisms designed to establish persistence in AWS and Azure, Aikido says.

Timeline

DateEvent
10 NovAikido Security says the wfx-loot operation began.
10-12 NovSafeDep records malicious commits in public repositories between 12:46 UTC on 10 November and 04:36 UTC on 12 November.
11 NovThe malicious version 3.14.3 of the Packagist package digitalelvis/neuronai-studio is published, according to its package listing.
12 NovMore probe pull requests from wfx-docs branches are opened between 05:04 and 05:38 UTC after Aikido's GitHub search ended; it is unclear whether the stealer followed.

How the campaign worked

Aikido says the probes went beyond repositories owned by the compromised accounts: 10,431 pull requests arrived through collaborator or organisation access, and 92% of those were still open when it checked. SafeDep found the stealer added to an existing workflow in 7,309 of the 8,093 repositories it counted, and as an npm postinstall hook in 771. The attacker usually reverted the change within minutes; where the stealer ran, logs were deleted, usually with the run itself.

StepSecurity says the attacker did not exploit a flaw in GitHub Actions; it used valid tokens belonging to developer and bot accounts. It traced 173 probe pull requests to tokens issued to the Fly.io GitHub app, could not determine how the tokens were obtained, and found no evidence that Fly.io's infrastructure was breached. Aikido's analysis does not explain how the 4,837 account credentials were obtained.

What the stealer took and left behind

Aikido says the stealer read GitHub Actions secrets and variables, workflow and OIDC tokens, AI service keys and credentials in runner process memory. It collected AWS, Azure, Google Cloud, Kubernetes, npm, PyPI and SSH credential files, and copied the repository, its last 100 commits and recent artefacts. SafeDep researchers found it sent data by HTTP POST to listed addresses or to Telegram; in a sample of 200 payload commits, about a fifth used Telegram only.

It exchanged GitHub OIDC tokens for AWS, Google Cloud and Azure sessions, and on AWS could create long-lived access keys. It could rewrite an AWS role trust policy to admit any AWS account and add Microsoft Entra ID federated credentials prefixed wfx-ci-, allowing workflows to sign in without a stored secret. Researchers list matching CloudTrail, GCP and Vault activity as signs to check.

Aikido says at least one Packagist package and one Go module were reached through automated publishing triggers. The malicious Packagist release digitalelvis/neuronai-studio 3.14.3 was still listed on 12 November under a malware flag; two versions of github.com/onuragtas/docker-env, v1.0.278 and v1.0.280, also carried the stealer. Aikido says at least 50 compromised users appear to publish packages, but assesses the attacker wanted CI secrets rather than package spread. It saw similarities with October's GhostAction campaign but judged the actors probably different.

What is not known yet

  • How the attacker obtained credentials for 4,837 GitHub accounts.
  • How many private repositories were affected and whether the stealer took data from every repository where it ran.
  • Who is responsible for the campaign; Aikido named no actor and judged it probably distinct from GhostAction.
  • Whether Fly.io or its GitHub app was breached; StepSecurity found no evidence and did not establish where the tokens came from.
  • Whether GitHub has confirmed the campaign or issued its own advisory.

What to do

If you own or write to a GitHub repository

  1. Search from 10 November for the wfx-docs, chore/ and ci-check branches, pull requests titled docs: dependency audit notes, and the commit messages listed in Aikido's and SafeDep's analyses.
  2. If any appear, treat every secret, variable and cloud role available to the repository's workflows as compromised, and treat any account that opened a probe pull request as compromised even if the stealer did not run.
  3. Revoke the commit author's tokens, OAuth app grants, SSH keys and sessions, enable 2FA, and rotate every secret the repository can read.
  4. Check AWS CloudTrail, Entra ID, GCP and Vault for the persistence signs the researchers list, including wfx-* sessions, CreateAccessKey, UpdateAssumeRolePolicy and wfx-ci-* federated credentials.

If you use the affected packages

  1. Block digitalelvis/neuronai-studio 3.14.3 and pin 3.14.2 or earlier.
  2. Treat any machine that installed or opened that version, or ran npm install on an affected commit, as compromised and rotate its keys.
  3. Treat github.com/onuragtas/docker-env v1.0.278 and v1.0.280 as malicious and stop using them.

For everyone else

  1. Limit repository write access and cloud role permissions to what each CI workflow needs, and review which bots and third-party apps can write to repositories.
  2. Monitor workflow file changes and cloud role assumptions for unexpected activity.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.