Research brief
GitHub reported in January 2025 that its researchers had secured more than 75 open-source GitHub Actions workflows and disclosed more than 90 vulnerabilities, then added workflow-focused CodeQL support. These figures describe the set its team examined and disclosed, not a prevalence rate for all repositories. A CI workflow sits in a privileged position: it may see source code, release permissions and deployment credentials while processing code or metadata from outside contributors. The development made workflow configuration a subject for security analysis rather than mere automation housekeeping.
The trust boundary
Events such as pull requests can carry attacker-controlled text or code into a workflow. The dangerous condition depends on how a repository combines event triggers, permissions and scripts. A job that needs only to run tests should not automatically inherit permission to publish packages or alter releases. Equally, third-party actions introduce a dependency with its own release and maintenance history. GitHub's article focused on finding risky patterns and strengthening workflows; it did not report that every one of the studied weaknesses had been exploited in the wild.
Defensive reading
Review workflow files like production code. Identify which events can start each job, the identity and permissions available to it, and whether untrusted inputs reach shell commands or privileged steps. Limit token permissions to the job's task, isolate release jobs from contribution checks, and review third-party action references and updates. Add static checks to catch regressions, but test the assumptions behind a rule before treating a green result as a security guarantee. When a workflow changes, ask whether a new path now connects outside input to secrets.
What remains bounded
The GitHub page has an initial January 2025 date and a later update; this entry uses the initial publication date and does not pretend to reconstruct an immutable first version. Its research sample is not a census. The broader lesson is that software provenance includes the automation that builds and ships an artifact. A repository can protect its source branch yet leak release authority through an overpowered workflow.