Tor Hides Your IP, Not Your Words: Free LLMs as an Attribute-Inference Threat to Users of Network Anonymity Tools

Tor and VPNs hide a user's IP address, and users often treat that as anonymity. We show that the text such users post remains a strong source of personal information, and that extracting it is now free. On 100 synthetic authors from the SynthPAI dataset, two free-tier Gemini models inferred eight personal attributes from the authors' comments. The best model guessed sex correctly 85.4% of the time, occupation 71.4% and country 72.3%, with a mean of 64.3% over all attributes; most of these attributes are never revealed by an IP address at all. A simple defense, rewriting comments with an LLM before posting, lowered mean accuracy to 49.5% but did not remove any attribute. We release all code and raw model outputs.