A new arXiv preprint, PrivDrift: Auditing User-Secret Leakage Under Topic Drift in Active LLM Conversations, addresses a growing concern: large language models are increasingly used as persistent assistants in shared and tool-augmented sessions. When users disclose sensitive information during an active conversation, that information may remain in the model's context and become vulnerable as the conversation shifts topics.
The paper proposes PrivDrift, an auditing approach specifically designed to detect and measure such leakage under topic drift. According to the abstract, the work targets user-facing, shared-session, and tool-augmented settings, where the risk of unintended secret exposure is heightened by the continuity of the conversation.
Because the source abstract is truncated, the full methodology and any experimental results are not described in the available text. The preprint is listed as arXiv:2609.30094, and readers interested in implementation details will need to consult the full paper.