Wiz's autonomous Red Agent walked into Snowflake's public snowflake-connector-net repo and turned a GitHub Actions flaw into a working exploit, no human in the loop. The jira_issue.yml workflow interpolated ${{ github.event.issue.title }} straight into a shell block, so any unauthenticated user could run arbitrary commands on a runner just by opening an issue.
The agent didn't just find the bug. It crafted an exploit, survived a failed first attempt, adapted the payload, and exfiltrated Jira credentials before Snowflake patched the same day. The workflow arrived via PR #1218, commit 4a1b8ce, merged June 18. Disclosure came June 23, with no CVE reference.
The script-injection vector is not new, and it's the part you can act on. Anywhere you interpolate ${{ github.event.* }} into a shell block, you've opened the same hole. Treat untrusted event fields as untrusted input, and if a runner touches secrets, assume an issue title is a payload.
The second lesson is attribution. GitHub pushed back on Wiz's claim that Copilot Autofix wrote the vulnerable code, pointing to a human-authored commit, even though Copilot reviewed the merged PR and flagged it all-clear. PR co-author metadata is not reliable enough to pin a vulnerable line on AI versus a human.