Open-Weight Decision Models Screen Web Passages Against Injection Attacks
Original title: Decision Models as Retrieval Gates Against Injection and Pollution
A Kev-4B gate reduces attack success from 27.9% to 5.7% on a 1,000-question benchmark.
The models read HTML passages to screen for injected instructions and synthetic pages. Attack success falls from 27.9% with no defence to 5.7% with the gate added. The gate takes 76 ms per passage and wrongly flags 4 of 186 benign pages.
- Attack success reduced from 27.9% to 5.7%.
- Takes 76 ms per passage.
- Wrongly flags 4 of 186 benign pages.
- Holds accuracy at 83% on pollution testbed.
From the original article
1 min read at arxiv.orgRetrieved web text can carry instructions aimed at the language model, or synthetic pages written to mislead. We test whether released open-weight decision models, which answer typed questions about a text without generating tokens, can screen retrieved passages against both threats.
By Mahdi Ghadamyari · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.