Reproduce “Reverse CAPTCHA: Evaluating LLM Susceptibility to Invisible Unicode Instruction Injection” 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

Marcus Graves

The authors test whether models follow answers hidden in invisible Unicode, and how tools and decoding hints affect compliance. They encode alternate trivia answers in two schemes and evaluate five models with and without Python execution. Haiku 4.5’s compliance rises from 0.8% without tools to 49.2% with tools, but unhinted tool-enabled rates remain 0–11%, so passive discovery is not established.

Summary by alphaXiv.

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