Staab et al. (ICLR 2024, peer-reviewed) showed language models infer location, income, and sex from plain text at up to 85% top-1. A 2026 preprint (Lermen et al.) ran it as an agent and matched 67% of Hacker News users to their real LinkedIn at 90% precision, for 1 to 4 dollars per person.
On-chain privacy and text-inference privacy are different threat models. CoinJoin, Silent Payments, and Monero protect the transaction graph. They do nothing for the forum posts, support requests, and replies that link your pseudonym to you. If you run a Bitcoin pseudonym, the text OPSEC is the half the privacy-coin conversation usually skips.
The post walks the three-stage attack chain (extract, search, verify) and the compartmentation playbook that breaks it.
Staab et al. (ICLR 2024, peer-reviewed) showed language models infer location, income, and sex from plain text at up to 85% top-1. A 2026 preprint (Lermen et al.) ran it as an agent and matched 67% of Hacker News users to their real LinkedIn at 90% precision, for 1 to 4 dollars per person.
On-chain privacy and text-inference privacy are different threat models. CoinJoin, Silent Payments, and Monero protect the transaction graph. They do nothing for the forum posts, support requests, and replies that link your pseudonym to you. If you run a Bitcoin pseudonym, the text OPSEC is the half the privacy-coin conversation usually skips.
The post walks the three-stage attack chain (extract, search, verify) and the compartmentation playbook that breaks it.