I spent a day going through the GitHub Actions workflows of some of the better known open source AI agent projects: AutoGen, Semantic Kernel, LangChain, Letta, Google's ADK, promptflow, smolagents, mem0, griptape, and a few others.
Nothing exotic. Just the boring stuff that gets skipped once a project grows fast: unpinned actions, CI checks that quietly stopped meaning anything, and a couple of straight up regressions where a fix from months ago got undone by a later commit nobody noticed.
A few examples worth mentioning, since the PRs are public.
Microsoft's autogen had a CI check meant to catch out of sync generated proto files. It could never fail. Someone had copied a demo snippet from a GitHub Action's README two years ago, and that snippet never included an exit code. The job kept reporting green no matter how stale the generated files got.
Google's adk-python had a fork guard on a pull_request_target workflow that was correctly fixed in an earlier PR, then quietly reverted by a later sync commit. That's scarier than a bug that was never fixed, because it means the protection existed and then silently disappeared.
griptape had an action pinned to a mutable tag instead of a commit hash, tj-actions/changed-files at v47. That's the exact action that got compromised in March 2025, when its version tags were rewritten to point at a commit that dumped CI runner memory, secrets included, into public workflow logs. Pinning to a tag instead of a hash means trusting the tag never moves. It moved once already, for this specific action.
None of it is glamorous. No RCE, no data breach, nothing headline worthy. It's the unglamorous work of checking whether a project's CI setup actually does what its authors think it does. In about half the repos I looked at, it didn't, for at least one workflow.
PRs are open on autogen, semantic-kernel, letta-code, adk-python, smolagents, mem0, griptape and promptflow. Some are already reviewed, most are waiting on a maintainer to get to them.
If you maintain a project whose CI has run unchanged for a year or two, it's worth a second look. The bugs aren't hard to find. Nobody had gotten around to looking.
Top comments (0)