GitHub's new AI agents promise to save devs 29 hours a week, with a side of privileged containers.

GitHub has launched its Agentic Workflows into general availability, a system that integrates AI coding agents directly into GitHub Actions. These agents, which can be powered by Copilot or other models, are designed to handle tasks like triaging issues or analyzing CI failures. They start with read-only permissions, and any proposed code changes are vetted by a separate threat-detection job before a distinct job with write tokens applies them.

However, GitHub's own FAQ warns that these safety nets don't automatically extend to custom jobs a developer might add, meaning the real security depends on careful configuration. To make things more interesting, the setup for self-hosted runners uses a Docker-in-Docker sidecar that runs as a privileged container by default—a potential security hole sitting right next to an agent system built on the principle of least privilege. While the marketing touts an "estimated 29 hours a week" saved, the actual safeguards feel like an afterthought you have to manually bolt on.