Embedding Translation Without Paired Data: A Security Wake-Up Call

A new arXiv paper from researchers at Cornell and elsewhere introduces an unsupervised method to translate text embeddings between different models' vector spaces—no paired data, no encoders, no predefined matches. The work leans on the Platonic Representation Hypothesis, which posits that neural models converge on a shared universal representation of reality.

The method maps any embedding into a universal latent space, then out to the target model's space. Across model pairs with different architectures, parameter counts, and training data, the translations achieve high cosine similarity. In other words, the geometry of embeddings is more universal than we assumed—and that has a dark side.

Security Implications for Vector Databases

The authors demonstrate that an adversary with access only to embedding vectors can translate them into a space they control, then run classification and attribute inference to extract sensitive information about the underlying documents. This is a direct threat to the common practice of storing embeddings in vector databases as a proxy for the original text.

If embeddings can be translated across models without any supervision, then treating them as opaque, secure representations is no longer viable. The paper is a practical proof that the geometry leaks meaning.

Why This Matters for Engineers

For anyone building RAG pipelines, semantic search, or any system that stores embeddings, this is a reminder that embeddings are not encrypted blobs. They carry semantic content that can be decoded with the right tools.

The upside is that translation could enable cross-model interoperability—e.g., using embeddings from a small model with a larger model's retriever. But the security tradeoff is real and underappreciated.