A within-subjects online experiment compares three proposed interfaces, that enable users to request and receive feedback on the selective deletion of personal information in an LLM-based system, and derives design recommendations for responsible, privacy-respecting, and user-controlled LLM systems.
The explosive prompt is introduced, a conditional payload that stays dormant until an attacker-chosen trigger is met, in effect a training-free, inference-time backdoor planted in a single piece of retrieved content.
Justin Szczepaniak, Elad Feldman, N. Viner et al.· arXiv.org· 0 citations
When interacting with large language models (LLMs), users have limited visibility and control over how their personal information is retained, particularly after incorporation into model training. Although some deployed LLMs provide options to opt-out of training, these mechanisms do not offer guarantees about the dele...