Summary
- Aikido identified 4,837 compromised GitHub accounts used to probe 61,764 repositories in the wfx-loot campaign.
- Malicious workflows executed in more than 1,800 repositories, potentially exposing CI secrets, cloud credentials, and software publishing permissions.
- The attackers attempted to erase workflow evidence, meaning repository owners may need to investigate execution history and credentials even where malicious files are no longer present.
Compromised GitHub accounts have been used to distribute credential-stealing software through automated development workflows, with security researchers identifying malicious activity involving tens of thousands of repositories and more than 1,800 instances of payload execution.
Belgian cybersecurity company Aikido Security identified the campaign, which it calls wfx-loot, after attackers began using stolen GitHub credentials on 10 November. The operation targeted secrets accessible within GitHub Actions, potentially exposing cloud infrastructure, application services, and software publishing systems.
In research published on 12 November, Aikido reported that 4,837 compromised accounts had been used to open probing pull requests across 61,764 repositories. Its investigation found evidence of malicious workflow execution in more than 1,800 repositories, although the full extent of credential theft and subsequent misuse remains uncertain.
The figures distinguish repositories approached by the attackers from those in which malicious code executed. A probing attempt does not establish that information was stolen from the target environment, whereas successful execution can expose credentials available to its automated workflows.
Compromised accounts opened the route into automation
The attackers used credentials belonging to existing GitHub users to interact with repositories those accounts could modify. This included projects owned by the compromised users and repositories accessed through organisational membership or collaboration arrangements.
Aikido identified more than 10,400 probing pull requests submitted to repositories outside the account holder’s own projects. The activity demonstrates how a stolen developer identity can affect organisations beyond the original victim, particularly where individuals contribute to multiple projects.
The initial requests were presented as routine documentation changes. After establishing write access, the attackers attempted to introduce malicious workflow modifications through several methods, including changes to existing build processes and temporary workflow files.
GitHub Actions is commonly used to automate software development tasks such as testing code, building applications, preparing packages, and deploying services. These activities can require access to credentials that are unavailable to ordinary users of the finished software.
A build workflow may authenticate with a cloud provider, retrieve protected dependencies, publish a software package, or deploy an application. The permissions granted to that workflow consequently determine how much additional infrastructure becomes accessible if the automation environment is compromised.
Aikido found that the malicious code attempted to collect GitHub Actions secrets, runtime credentials, cloud access information, and authentication material associated with development tools. Its capabilities also extended to credentials used by AI services, container systems, and software registries.
The payload contained functions intended to expand access where the available permissions allowed it. Those capabilities did not establish that every compromised workflow could perform the same actions, but they increased the potential consequences in environments where automation possessed broad authority.
Short-lived changes complicated detection
Once the malicious workflow had been introduced, the attackers attempted to reverse their modifications and delete the associated execution logs. Aikido observed cases in which this activity occurred within a minute of the initial change.
That behaviour can leave a repository appearing substantially unchanged during a later inspection. A developer reviewing the current files may find no malicious workflow, even though an earlier version executed and obtained access to sensitive information.
Researchers identified missing workflow run records as one possible indication of activity. However, the absence of visible malicious files or logs cannot independently establish that a particular environment was unaffected.
Aikido’s initial report contains slightly different execution counts in its summary and detailed analysis. Both indicate activity affecting more than 1,800 repositories, which is the figure used here rather than treating one preliminary count as the definitive total.
The investigation also found that many probing requests remained open when the researchers collected their data. That observation describes the campaign at the time of analysis and should not be interpreted as a current measurement of unresolved incidents.
Organisations investigating possible exposure may need to examine repository activity, available audit records, credentials, and downstream systems rather than relying exclusively on the present state of their source code.
Publishing systems introduce additional exposure
The campaign reached some published software packages through automated release processes, including identified versions of a Packagist package and a Go module.
The consequences differed between them. Aikido warned that the affected Packagist package could expose systems interacting with the malicious version, while the identified Go module versions contained the payload within GitHub Actions workflow files rather than the Go source code or compiled release binaries.
Consequently, using one of those Go module versions did not automatically mean that a downstream application executed the credential stealer. The distinction depends on whether the malicious content was actually run within the affected development environment.
Nevertheless, access to software publishing systems increases the potential reach of an initial compromise. Credentials obtained from a build environment can provide authority to publish packages or alter release processes, creating risks for organisations that depend on the resulting software.
Aikido also identified overlap between accounts involved in wfx-loot and an earlier campaign called GhostAction. The researchers assessed them as distinct operations, noting differences in infrastructure and techniques despite some shared compromised accounts.
The presence of the same victim in multiple incidents does not establish that the same attacker was responsible. Stolen credentials can circulate between different criminal operations, exposing accounts to repeated misuse.
Recovery depends on the permissions exposed
Removing the malicious workflow does not necessarily complete an organisation’s response. Where the payload executed, credentials accessible to the workflow may need to be treated as compromised and replaced.
Security teams must also examine whether cloud permissions, application identities, or publishing arrangements were changed during the incident. Replacing an exposed token may not remove access obtained through another authentication mechanism or a newly created cloud identity.
The wider investigation therefore depends on understanding what the workflow was authorised to access when the malicious code ran. A repository restricted to isolated testing presents different consequences from one holding credentials capable of modifying production infrastructure.
Narrowly scoped permissions and separation between build, testing, and deployment activities can limit the consequences of such incidents. Organisations may also need controls governing who can change workflow files and whether sensitive credentials are available to automation triggered by repository modifications.
Those measures must reflect the operation of individual development environments. A workflow legitimately requiring deployment permissions cannot simply be stripped of all access, but its authority can be limited to the services and activities necessary for its function.
Aikido’s findings establish a large campaign involving compromised developer accounts and observed workflow execution. They do not establish the total number of affected businesses, the value of information stolen, or the extent of subsequent attacks using the credentials.
For organisations using GitHub Actions, the immediate risk depends on whether their repositories were affected and what secrets their workflows could access. The investigation illustrates how developer identities and automated permissions can provide a route from source code management into wider business infrastructure.












