PrivDrift is a for auditing whether user-disclosed secrets remain recoverable after topic drift in active LLM conversations. The authors constructed 1,000 multi-turn dialogues with seeded secrets, content-dense drift turns, and standardized extraction probes, and evaluated three LLMs.
Across models, dialogue-level hybrid leakage ranged from 38.7% to 54.6%. Persuasion intensity and secret type significantly affected leakage, while additional topic drift (up to six turns) did not reliably reduce it.
These findings highlight that privacy risk in active conversations is a persistent behavioral failure, not limited to training-data extraction or immediate jailbreaking. The work underscores the need for explicit context privacy mechanisms.