Reproduce “Decision Hijacking: Prompt Injection Attacks on Jev's Typed Probabilistic Decisions” 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
Tiantong Wu, Wei Yang Bryan Lim
The authors test whether indirect prompt injections can steer Jev, a constrained decision model, among declared actions. They reconstruct 510 direct-harm cases, alter only observation content, and use a deterministic search guided by Jev’s reported probabilities. The base attack selects its target in 1.8% of cases; adaptive feedback raises validated ASR to 3.5%, in one model version and benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.