Reproduce “GitInject: Real-World Prompt Injection Attacks in AI-Powered CI/CD Pipelines” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jafar Isbarov, Umid Suleymanov, Ilia Shumailov, Murat Kantarcioglu

The authors introduce GitInject, which evaluates prompt injection by running CI/CD workflows on real GitHub runners rather than relying only on simulated agent environments. In tested configurations, unmerged pull-request files loaded as trusted provider guidance enabled attacks, and AgentDojo marked 10 of 14 live successes safe (71.4%). These findings cover only the named attacks and workflows tested.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.