Attention layers leak private data in tabular foundation models

arXiv·medium signal

arXiv 2606.26021 (cs.CR) challenges the assumption that tabular foundation models are low-privacy-risk because they pretrain on synthetic data, showing attention layers expose privacy vulnerabilities and proposing protection for high-risk queries. As agents increasingly call tabular FMs over enterprise data, this is a concrete data-governance concern. A security-focused result builders should weigh before treating synthetic-pretrained tabular models as safe by default.