DNA foundation models --- DNABERT-2, Evo 2, Nucleotide Transformer v2 --- advertise embeddings as privacy-preserving representations of genomic sequences, but decoder networks fed zero-shot per-token embeddings achieve near-perfect sequence reconstruction across all three models. Mean-pooled embeddings degrade with sequence length but substantially exceed random baselines. Evo 2 and NTv2 are most vulnerable, with reconstruction similarities exceeding 90% for short sequences, while DNABERT-2's byte-pair encoding tokenization provides partial protection. The fundamental vulnerability traces to a correlation between embedding similarity and sequence similarity --- the very property that makes embeddings useful for downstream tasks is precisely what makes them invertible.
Embeddings-as-a-Service platforms rest on an implicit assumption that dense vector representations obscure their inputs. For natural language, this assumption has already been challenged; for genomic data --- where the sequences are finite-alphabet, the stakes are medical, and the identifiability is permanent --- the failure is more consequential. The attack does not require access to the model architecture or weights, only to the embeddings themselves. BPE tokenization's partial resistance is informative: it breaks the one-to-one mapping between positions and tokens, introducing a combinatorial barrier that character-level and k-mer tokenizations lack. The protection is accidental --- a side effect of compression, not a privacy design.
Note (2026-11-15; read against the paper's full text): the essay says the attack does not require access to the model architecture or weights, only to the embeddings. In the paper's pipeline the decoder is trained on 100,000 sequences per configuration whose embeddings are computed with each foundation model (DNABERT-2, Evo 2, Nucleotide Transformer v2), and the text discusses white-box and black-box access only as a general taxonomy of inversion attacks; I found no statement that the attack works without being able to run the model.
The utility-privacy tension in representation learning is not a tradeoff to be optimized but a structural constraint: any embedding that preserves task-relevant information preserves reconstruction-relevant information, because the two are the same information viewed from different directions. The only embeddings that are truly private are the ones that have destroyed the signal --- and those are useless. Every claim that a transformation “preserves utility while ensuring privacy” must specify exactly which dimensions of utility are orthogonal to which dimensions of identity, and for genomic data, that orthogonality does not exist.
(arXiv:2603.06950)