Aikido Security has identified a vulnerability in GitHub Actions and GitLab CI/CD workflows that allows AI agents to execute malicious instructions, potentially leaking sensitive information. The flaw affects multiple companies and demonstrates how AI prompt injection can compromise software supply chains.

Grafana Labs introduced Zizmor, an open source static analysis tool, in their CI/CD pipelines to detect and prevent vulnerabilities in GitHub Actions following a security incident. The tool helps identify unsafe configurations and practices, such as the use of `pull_request_target`, and is part of a broader effort to enhance security across their repositories. Despite facing challenges like GitHub's rate limiting, Grafana is committed to using Zizmor to bolster their defenses against future attacks.

The article discusses the security considerations necessary for using GitHub Actions in CI/CD setups, emphasizing the importance of protecting workflows against potential threats from contributors with write access. It details various attack scenarios, including script injection vulnerabilities, and provides best practices for securing sensitive workflows and managing permissions effectively.

Sysdig's Threat Research Team uncovered significant security vulnerabilities in GitHub Actions workflows across popular open source projects, including those by MITRE and Splunk. Their research revealed how insecure configurations, particularly using pull_request_target, can expose sensitive credentials and allow for exploitation, prompting the team to recommend best practices to enhance CI/CD security.