Openflows

Signal

Large-scale online deanonymization with LLMs

The paper [Large-scale online deanonymization with LLMs] reports that LLM-based pipelines can match pseudonymous online identities across datasets using raw text alone, with strong precision-recall performance against classical baselines.

Context

The core shift is methodological. Earlier deanonymization approaches often depended on structured data and hand-engineered features; this work uses model-assisted feature extraction, candidate retrieval, and verification directly on unstructured content.

Relevance

For Openflows, this is a high-importance privacy and governance current. If pseudonymous traces can be linked at scale, communication safety, civic organizing, and platform trust models need stronger default protections.

Current State

Newly published research signal (submitted February 18, 2026) with immediate threat-model implications.

Open Questions

• Which practical defenses most reduce cross-platform linkability without collapsing usability?
• How should platforms update privacy guidance for users relying on pseudonymity?
• What evaluation standards should separate legitimate identity resolution from abusive deanonymization tooling?

Connections

• Linked to pseudonymity-collapse-response as the implications circuit.
• Linked to signal-org as directly affected communication infrastructure.

Updates

2026-03-23: The paper was revised to version 2 on February 25, 2026, specifying performance metrics of up to 68% recall at 90% precision across datasets including cross-platform linking between Hacker News and LinkedIn. This update concretizes the threat model with empirical data rather than general claims of strong performance.