Skip to content
arXiv· arXiv:2605.10614· Riya Tapwal, Abhishek Kumar, Carsten Maple·Indexed by this site · Originally published

PRISM: Generation-Time Detection and Mitigation of Confidential Leaks in Multi-Agent LLM Pipelines

PRISM: Generation-Time Detection and Mitigation of Secret Leakage in Multi-Agent LLM Pipelines

AI summary

PRISM is a generation-time leak prevention mechanism for multi-agent LLM pipelines, achieving a 0.0% task-level leak rate on a 2,000-task benchmark. The decoding step fuses 16 types of signals and intervenes token by token across green, yellow, and red zones, achieving F1=0.832, precision=1.000, and recall=0.712 across 13 types of attacks, outperforming Span Tagger's F1=0.719 and 15.0% task-level leak rate.

Read originalarxiv.org