Poisoned Pipeline Execution

What the technique is

Adversaries may manipulate continuous integration / continuous development (CI/CD) processes by injecting malicious code into the build process. There are several mechanisms for poisoning pipelines: In a Direct Pipeline Execution scenario, the threat actor directly modifies the CI configuration file (e.g., gitlab-ci.yml in GitLab). They may include a command to exfiltrate credentials leveraged in the build process to a remote server, or to export them as a workflow artifact. In an Indirect Pipeline Execution scenario, the threat actor injects malicious code into files referenced by the CI configuration file. These may include makefiles, scripts, unit tests, and linters. In a Public Pipeline Execution scenario, the threat actor does not have direct access to the repository but instead creates a malicious pull request from a fork that triggers a part of the CI/CD pipeline. For example, in GitHub Actions, the pull_request_target trigger allows workflows running from forked repositories to access secrets. If this trigger is combined with an explicit pull request checkout and a location for a threat actor to insert malicious code (e.g., an npm build command), a threat actor may be able to leak pipeline credentials. Similarly, threat actors may craft pull requests with malicious inputs (such as branch names) if the build pipeline treats those inputs as trusted. […]

What MITRE says you would watch

• AN1473

  Detects anomalous CI/CD workflow execution originating from forked repositories, with pull request (PR) metadata or commit messages containing suspicious patterns (e.g., encoded payloads), coupled with the use of insecure pipeline triggers like pull_request_target or excessive API usage of CI/CD secrets. Correlation with unusual artifact generation or secret exfiltration via encoded or external network destination URLs confirms suspicious behavior.