Decision Models as Retrieval Gates Against Injection and Pollution

Summary

A study tests open-weight decision models as gates to screen retrieved web passages against prompt injection and content pollution. On a 1,000-question benchmark, a Kev-4B gate reduced attack success from 27.9% (no defense) to 5.7% (with published mitigations), taking 76 ms per passage and flagging 4 of 186 benign pages. On a 200-query pollution testbed, filtering before top-ten selection kept fake pages out (0.5% exposure) and held accuracy at 83%, versus 28% with no filter.

This is a reviewed summary. Read the complete article at the original source.

Why it matters by industry

• Artificial intelligence

  This research introduces a lightweight, open-weight method to improve retrieval safety in AI systems, relevant to models, agents, and open-source development.