Researchers propose Prism for OOXML extractor divergence attacks

Researchers demonstrate that Office document extractors used by large language models often interpret OOXML files differently from what human reviewers see, introducing security vulnerabilities. By analyzing widely used extractors, the authors identified multiple channels where hidden content is read or visible content is dropped, manipulated, or reordered without changing the visual display. To address these extraction divergences, they propose Prism, a lightweight defense that reconstructs the human view prior to model processing. Open article →