Reproduce “Agentic Backdoors via Pretraining Poisoning” 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
Xinyan Hu, Pingbang Hu, Chloe Loughridge, Xuanli He, Nicholas Carlini, Matthew Jagielski, Javier Rando, Yiming Zhang
The authors test whether pretraining-poisoned trigger–command rules persist through clean training for command-line agents. They plant associations in varied prose, then train Qwen3 models through supervised fine-tuning, preference optimization, and GRPO. In Qwen3 4B, declaration poisoning reached 92.7% Pass@32 after GRPO, versus 3.2% for demonstrations; tests used synthetic paths and did not execute commands on real machines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.