Reproduce “DuplexJail: Spoken Interruption Attacks on Full-Duplex Speech Models” 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

Jaechul Roh, Deepak Chandran, Amir Houmansadr, Andrea Fanelli

DuplexJail tests whether timed spoken interruptions through the user audio channel affect harmfulness in full-duplex speech models. It sends fixed prompts during or after responses. On AdvBench PersonaPlex reaches 40.3% whole-response ASR versus a 6.5% baseline. On HarmBench the prompt raises PersonaPlex’s ASR by 14.7 percentage points; effects vary across models and do not establish overlap as the cause.

Summary by alphaXiv.

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