Decoding the world of cybersecurity

Stolen GitHub accounts expose build pipeline credentials

Stolen GitHub credentials enabled widespread attacks against software development pipelines. Researchers identified thousands of compromised accounts probing more than 61,000 repositories and stealing secrets from automated workflows.

Stolen GitHub accounts expose build pipeline credentials
Summary
  • Aikido identified 4,837 compromised GitHub accounts used to probe 61,764 repositories during the wfx-loot campaign.
  • The malicious payload executed and exfiltrated credentials in at least 1,835 repositories, according to the researchers' published findings.
  • Attackers attempted to conceal workflow changes and establish persistence in connected cloud systems, potentially extending compromise beyond the original GitHub credentials.

Attackers have used thousands of compromised GitHub accounts to introduce credential-stealing malware into software development pipelines, with researchers identifying more than 61,000 repositories probed for access and evidence of malicious execution in at least 1,835.

Research published by Aikido Security on 12 November traces the campaign, named wfx-loot, to activity beginning on 10 November. The operators used stolen GitHub credentials to identify repositories they could modify before attempting to introduce malicious code into automated build and deployment processes.

Aikido identified 4,837 compromised accounts involved in opening probe pull requests across 61,764 repositories. Of those requests, 10,431 targeted repositories that the compromised account holders did not own, relying instead on existing collaborator or organisational permissions.

The scale of probing does not correspond directly to successful malware execution. Researchers identified at least 1,835 repositories where the payload ran and credentials were exfiltrated, while many other repositories received only an initial access test.

Legitimate accounts provided the initial access

The operators did not need to exploit a newly discovered GitHub platform vulnerability. They used valid credentials associated with developer and automated accounts, allowing them to interact with repositories according to the permissions already granted to those identities.

The first stage involved opening pull requests that appeared to concern routine documentation or dependency maintenance. Where an account possessed sufficient write permissions, the attackers then attempted to introduce malicious code through development workflows or installation processes.

GitHub Actions automates functions including software testing, package creation and application deployment. These workflows frequently interact with external services, using credentials to access cloud infrastructure, publish packages or retrieve information required during a build.

That access made the workflows an attractive target. The malicious code attempted to collect credentials exposed to the runner, including GitHub authentication tokens, cloud service credentials and keys associated with developer tools and AI platforms.

Depending on the permissions available, the payload could also seek access to Amazon Web Services, Microsoft Azure and Google Cloud. The resulting exposure could extend beyond a project’s source code to infrastructure managed through its automated development processes.

Not every identified capability was necessarily exercised in every affected environment. The researchers distinguished the functions implemented in the malware from what could be confirmed through the available repository and execution evidence.

Attackers attempted to erase workflow evidence

The campaign incorporated measures intended to conceal unauthorised changes after the malicious code had executed.

Aikido observed cases in which attackers reverted changes shortly after execution and deleted associated workflow runs and logs. The affected repository could consequently appear normal during a later inspection even though an unauthorised workflow had already accessed credentials.

Independent analysis by StepSecurity identified similar behaviour in more than 100 examined malicious commits. Its investigation also found changes in the malware’s communication methods during 12 November, including later samples that used Telegram rather than the original collection infrastructure.

The changing communication arrangements complicate detection based solely on previously identified network destinations. More fundamentally, the rapid reversal of repository changes makes historical audit records important to establishing whether a suspicious workflow executed.

A pull request used to test access is not itself evidence that the credential-stealing payload ran. However, such activity can establish that an account was used without authorisation and that repositories accessible through its existing permissions may require investigation.

The distinction becomes particularly significant where one compromised developer identity has access to projects belonging to several organisations.

Cloud access could persist after credentials change

The malware contained functionality intended to extend access beyond the original development environment.

Aikido identified attempts to obtain additional cloud permissions and alter identity configurations in ways that could survive the replacement of stolen credentials. These included changes to cloud role trust arrangements and the creation of new federated identity credentials.

Such persistence can undermine a response limited to revoking the original GitHub token. Even after the repository account is secured, an attacker may retain a separate route into connected infrastructure if additional cloud permissions were successfully established.

That outcome depends on the authority granted to the compromised workflow. An automated process without permission to change cloud identity configurations cannot necessarily perform those actions merely because its execution environment has been compromised.

Nevertheless, development pipelines commonly hold privileges that permit software deployment, infrastructure changes or access to production services. A workflow’s effective permissions therefore determine how far the consequences of credential theft can extend.

StepSecurity also identified instances in which the attackers modified existing scripts used during automated builds, expanding the possible locations in which malicious code might execute.

Software packages carried additional exposure

Aikido reported that the campaign reached at least one published Packagist package through automated release processes. It also identified affected Go module versions containing malicious workflow files.

Those findings represent different types of risk. Malicious code executed during package installation can affect systems consuming that package, while a repository containing a compromised workflow file does not necessarily cause the same code to run when its ordinary software is downloaded or built.

The researchers have also identified overlap between some compromised accounts and an earlier campaign known as GhostAction. They assessed the operations as distinct, however, citing differences in infrastructure and methods.

Reused stolen credentials could explain the overlap without establishing that the same operator controlled both campaigns. The identity of the actor responsible for wfx-loot remains unconfirmed.

The breadth of the attack reflects the reach of legitimate development permissions across organisational boundaries. A compromised account may be able to alter repositories owned by employers, clients, open-source projects or other collaborators, exposing multiple parties through one identity.

Where the credential-stealing payload executed, investigation must extend to the secrets accessible through the affected workflow and the cloud permissions those credentials could have obtained. Restoring repository files alone does not establish that connected systems remain secure.

Aikido’s 12 November findings establish at least 1,835 repositories with observed malicious execution and credential exfiltration. The total affected population remains uncertain, particularly where private repositories, deleted workflow records or later activity limit the researchers’ visibility.

×