A team led by Zhuchenyang Liu published a paper on Oct 8, 2026 showing that the vector databases used by visual document search systems can be inverted back into readable pages. The attack recovers about 47% of the words and ranks the original page first in 98.4% of queries on the ViDoRe v3 benchmark. Simple defenses like token pooling or shuffling reduce word recovery to roughly 8%, but a model that re‑orders shuffled vectors can still rank the correct page 93.5% of the time. The findings apply to another retriever as well, with source pages ranked first 70.2% of the time.
Why it matters
Organizations storing document indexes in third‑party vector databases may need stronger protections to keep confidential text safe.
