Git histories reveal how agent skills spread
In plain words
Agent skills are instruction files that coding agents can use. Because people copy them between repositories, a snapshot can show where a skill is present but not where it came from or whether later edits reached its copies.
This paper builds dated copy links from Git histories and uses them to rank repositories for review. In the reconstructed network, only 20.9% of copies followed a later edit to their source, suggesting that source changes often do not travel through copies.
The research
When a skill is copied between repositories, a change at its source may not reach the copies. Without knowing who copied from whom, it is hard to trace that spread or decide which repositories to review. The paper asks whether Git histories can recover those paths and help target review.
It reconstructs dated copy links from Git histories and uses a source-choice model to rank repositories for security review.
A copy’s source can change a skill, but does that change reach repositories holding copies? Without dated copy links, a snapshot cannot answer that question or show where review might prevent further spread.
The researchers reconstructed links from Git histories and found that only 20.9% of copies followed a later source edit. They also found that auditing the model’s top 100 repositories prevented more later flagged-skill adoptions than auditing the 100 most-starred.
The data cover public GitHub during the format’s first ten months. High-risk flags mark capabilities, not confirmed malice; because the flags miss some cases, the reported risk counts are lower bounds.
Analysis
Deep
- Authors
- Fahd Seddik
- License
- CC BY 4.0