TheMachine Press

A daily newspaper for the age of artificial intelligence.

Morning editionPermanent story

safety

The Leak Lived in the Pipeline, Not the Language Label

English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.

Published Updated Story ID: mp-2026-08-09-018
Read the complete editionStory JSON

Summary

English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.

A five-language audit of one Qwen-based RAG pipeline found English had the highest unstructured synthetic-PII leak rate with output filtering alone. Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost.

Why it matters

English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.

Limits and context

  • Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost.

Key claims

  1. English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.

    Qualification: Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost.

    Evidence: source-2026-08-09-020

Sources

  1. arXiv preprint 2608.05163arXiv · primary research

Corrections

No corrections have been recorded for this story.