Profiling for a Few Cents: Unveiling the Privacy Iceberg of LLM Agents

The researchers propose the PrivacyIceberg framework and develop IcebergExplorer, which uses a very small amount of personally identifiable information as search seeds to reconstruct high-fidelity personal profiles, achieving over 90% factual accuracy within 10 minutes in real-world scenarios at a cost below 3 US dollars. PrivacyIceberg classifies real-world privacy risks into three levels according to the complexity of LLM exploitation: explicit search, contextual inference, and deep aggregation. The researchers also examine privacy perception from both human cognition and the practices of LLM integration platforms, finding that platforms have not responded to public concerns in either technology or policy. Based on the audit results, they summarize six categories of causes leading to privacy leakage and propose multi-party countermeasures for LLM providers, individuals, and data publishers.