Reproduce “Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using 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
Farhan Zarif, Kazi Shahid Ahmed Galib, Synthia Akter Promi, Shabnam Sharmin Tarin, Ariyan Hossain, Sadif Ahmed, Farig Yousuf Sadeque
The authors audit indirect prompt injection outcomes across eleven open-weight agents on InjecAgent and AgentDojo, distinguishing ignored instructions, refusals, and other failures. Among benchmark-scored failed attacks, agents ignored injections in 76.3% and 80.3% of cases, while genuine refusals accounted for 18.0% and 17.2%. Prompt changes sharply shifted attack success; task-completion analysis was limited to AgentDojo.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.