Task-graph contextual integrity limits multi-agent privacy leaks
Original title: Judging Before Propagation: Task-Graph Contextual Integrity for Multi-Agent Privacy
On MAPLE-Bench's 144 cases, TG-CI cut mean unauthorized disclosures 85.8-94.1% while keeping 93.6-96.1% of baseline task accuracy.
Researchers formalize Task-Graph Contextual Integrity for multi-agent LLM workflows whose persistent memory and inter-agent communication can make an otherwise acceptable disclosure inappropriate downstream. TG-CI models a workflow as a Multi-Agent Task Graph and authorizes propagation with Full-Scope and Auto-Region policies. MAPLE-Bench has 144 cases that vary topology, disclosure history, compromised-agent placement, and item-level privacy policies. Against a Local-History CI baseline, the policies trade privacy, cost, and utility.
Unconfirmed: The authors say TG-CI is, to their knowledge, the first task-graph-native formalization; results are from their configuration.
- MAPLE-Bench contains 144 cases across four scenarios.
- Unauthorized disclosure counts fell 85.8-94.1% versus Local-History CI.
- Task accuracy retained was 93.6-96.1% of the baseline.
- Full-Scope checks the complete downstream scope; Auto-Region uses frontier gates.
From the original article
1 min read at arxiv.orgMulti-agent LLM systems increasingly rely on persistent memory and inter-agent communication. These capabilities create privacy risks that extend beyond individual interactions: a release appropriate in isolation may become inappropriate when combined with information previously disclosed to a recipient or propagated to downstream agents along the planned workflow.
By Hengyu Wu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.