Anthropic-ETH study sparks $500 bounty for LLM deanonymization attacks

A researcher has posted a $500 bounty for red-teamers capable of building attack pipelines that use LLMs to deanonymize online accounts. The bounty directly builds on a recent study from Anthropic and ETH Zurich, which demonstrated how combining stylistic analysis—sometimes called a writer's DNA—with contextual metadata like age, job, and location could reliably link anonymous Reddit accounts to their real-world identities. The researcher wants to test whether existing models can replicate this technique on their own public blog posts, offering an additional $500 if an anonymous Reddit account is successfully uncovered.

The request has ignited debate within the security community. Requirements for participation include proven red-teaming experience, strong knowledge of stylometry (the statistical analysis of writing style), and a pledge to keep findings confidential. While the bounty is small compared to corporate vulnerability rewards, it highlights a growing risk: as LLMs improve, the combination of linguistic fingerprinting and inference from public text could make anonymity increasingly fragile. Privacy advocates argue that even a token bounty normalizes dangerous surveillance techniques. The researcher insists the goal is to raise awareness, similar to how responsible disclosure works in security research, but the potential for misuse remains a central concern.