Reproduce “Competition Report: Finding Universal Jailbreak Backdoors in Aligned LLMs” 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
Javier Rando, Francesco Croce, Kryštof Mitka, Stepan Shabalin, Maksym Andriushchenko, Nicolas Flammarion, Florian Tramèr
The competition tests recovery of hidden suffixes planted in five LLaMA-2 7B models. Participants searched candidate strings with a safety reward model, while teams prioritized tokens with large embedding differences across models. Teams recovered some triggers exactly or nearly, but the best team’s aggregate score (−33.414) remained above the planted-trigger baseline (−37.733). These scores concern held-out prompts, not human judgments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.