This week’s issue looks at a growing problem: how easily “anonymous” data can now be re-identified using large language models. From browsing histories to research interviews and pseudonymous accounts, studies show that LLMs can connect small clues across the web to reveal real identities.

#### One YouTube Video

In 2016, a data scientist and an investigative journalist set up a fake marketing company, bought 3 billion URLs from 3 million German users, and de-anonymised them in days. Today, anyone with access to an LLM can do this even faster.

#### One Research Paper

Anthropic published anonymised interview transcripts from 1,250 professionals. In this paper, a Northeastern researcher showed that an LLM agent with web search could re-identify participants by matching details in their answers to published academic papers—no specialist skills required.

[**Read more**](https://arxiv.org/pdf/2601.05918)

#### One Article

Researchers show how LLM agents can link pseudonymous accounts across Hacker News, Reddit, and LinkedIn. Using embedding-based search and reasoning, the system identifies users with high precision at a large scale.

[**Read more**](https://simonlermen.substack.com/p/large-scale-online-deanonymization)

#### One Infographic

Source: [Substack](https://simonlermen.substack.com/p/large-scale-online-deanonymization)

#### One Way to Mitigate Risk

The research above shows why anonymising data when publishing statistics is no longer enough. A stronger alternative is differentially private synthetic data, a dataset that preserves the statistical properties of the original without exposing the individuals behind it.

[**Read more**](https://www.nist.gov/blogs/cybersecurity-insights/differentially-private-synthetic-data)

#### One Meme

Source: [PostgreSQL](https://www.postgresql.eu/events/fosdem2019/sessions/session/2287/slides/151/postgresql_anonymizer.reveal..pdf)