Study shows LLMs crack pseudonymity

This is the kind of paper that turns a privacy assumption into a deprecated security model.

• –The real leap is unstructured text: attackers no longer need neatly linked datasets when LLMs can extract identity clues, search the web, and verify candidates on their own.
• –The risk is bigger than doxxing; the same workflow can support hyper-targeted phishing, corporate profiling, and large-scale surveillance of critics or vulnerable users.
• –Platforms that rely on public posting histories now have a stronger case for aggressive anti-scraping controls, API rate limits, and data-minimization defaults.
• –LLM vendors will face growing pressure to treat deanonymization as a first-class abuse category, not just a side effect of general web-enabled agents.