Reproduce “Kerkoporta: Injecting Structured-Output Backdoors in Reasoning Coding Agents” 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

Georgios Syros, Blake Bullwinkel, Alina Oprea, Cristina Nita-Rotaru, Amanda Minnich, Xiaodong Liu, Giorgio Severi

Kerkoporta demonstrates how a poisoned coding-agent model can turn a trigger in a tool result into an executable, schema-valid tool call. The authors use supervised fine-tuning to teach this behavior, then online reinforcement learning to recover clean-task performance. On Qwen3.5-4B, six RL steps bring timeout-excluded rewards near clean baselines while attack success remains similar in tested models, harnesses, and sandboxed trials.

Summary by alphaXiv.

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