Reproduce “Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI 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
Ilan Zendel, Matan Levy, Stav Cohen, Avi Mendelson, Amit LeVi
The authors test whether fabricated safety claims in public reviews can steer AI agents toward an attacker’s listing. In controlled recommendation tasks, safety payloads succeed on 60.4% of queries, versus 8.2% for misconduct complaints and 6.8% for benign-positive reviews. A corroboration instruction lowers success to 3.6%, while the hazard remains visible in 98.4% of answers, within the tested non-adaptive setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.