Large-scale online deanonymization with LLM agents
a day ago
- #LLM Security
- #Deanonymization
- #AI Privacy
- LLMs can deanonymize users from anonymous online posts with high precision.
- Deanonymization scales to tens of thousands of candidates and is increasingly practical.
- LLMs infer personal details (location, job, interests) from a few comments and search the web.
- AI surveillance and deanonymization pose concrete harms, enabling spear-phishing and exploits.
- Two proxy benchmarks used: cross-platform matching and matching split accounts.
- Platforms can mitigate risks by restricting data access and enforcing rate limits.
- LLM providers face challenges with refusal guardrails and monitoring due to task decomposition.
- Individuals should adopt a stronger privacy mindset, as shared details create unique fingerprints.
- Open-source models lack safety guardrails, making misuse harder to prevent.