LLMs can identify anonymous forum users with 68% accuracy at 90% precision
How the de-anonymization works
A research team gathered thousands of posts from anonymous forums like Hacker News and Reddit, then asked language models to identify the authors. They used Hacker News profiles connected to LinkedIn as ground truth, anonymized them, and fed them to AI systems.
The AI was given prompts like: "Which candidate is the same person as the query? Consider overlapping traits like location, profession, hobbies, demographics, and values. A match should share multiple distinctive traits, not just one or two common ones."
Key findings from the study
- Models identified 68% of anonymous users with 90% precision
- This compares to "near 0% for the best non-LLM method"
- Gemini and ChatGPT completed the task in minutes versus hours for humans
- The research shows "practical obscurity protecting pseudonymous users online no longer holds"
What AI can extract from anonymous posts
The models don't just look for explicitly stated personal details. Researchers provided examples of what can be inferred from years of comments:
- Location (Nelson, British Columbia, Canada)
- Profession (pediatric nurse)
- Demographics (woman, married, two daughters)
- Possessions (owns a Prius)
- Hobbies (plays Stardew Valley, fan of Critical Role)
- Preferences (supports nuclear energy, celiac, does not like cilantro)
- Behavioral patterns (visits Berlin subreddit, uses British spelling, accidentally wrote a "¿" in English text)
Implications for online privacy
According to researcher Daniel Paleka from ETH Zurich: "People sometimes express their opinions through pseudonymous accounts, assuming that those opinions will remain private. The existence of a mechanism to investigate or monitor with large language models that allows us to simply ask about a person's beliefs, political opinions, insecurities, or anything else that can be extracted from their anonymous Reddit account, for example, could disempower many people today."
Paleka notes that models can provide a timeline of a person's life if there's sufficient information online, and warns: "Keep in mind that everything you post stays on the internet and can become the target of future models" that will be even more effective.
📖 Read the full source: HN LLM Tools
👀 See Also
MCP Package Security Scan Reveals Widespread Destructive Capabilities Without Confirmation
A security scan of 2,386 MCP packages on npm found 63.5% expose destructive operations like file deletion and database drops without requiring human confirmation. The researcher discovered 49% had security issues overall, with 402 critical and 240 high severity vulnerabilities.
Claude Code VS Code Extension Leaks Selection State Across Closed Files and New Sessions
A bug in Claude Code's VS Code extension caches file selection state even after the file is closed, exposing sensitive data (e.g., Supabase service-role keys) to a brand new CLI session. Full repro steps and GitHub issue #58886.
Developer Builds Firecracker MicroVM Sandbox for OpenClaw Security
A developer concerned about LLM security built a bare-metal sandbox using Firecracker microVMs to isolate OpenClaw scripts, with each script running in its own Linux kernel with 128MB RAM cap and no network by default.
Analysis of Claude Code's Instrumentation and Telemetry Capabilities
A source code analysis reveals Claude Code implements extensive behavior tracking including keyword-based sentiment classification, permission prompt hesitation monitoring, and detailed environment fingerprinting.