Answers are grounded in this paper and cite their sources.
Multi-agent systems using the Model Context Protocol (MCP) can lose provenance across storage and session boundaries. This spatio-temporal data loss enables an adversary on the ingestion path to distribute an outbound-transfer directive across incomplete records for a privileged agent to assemble and execute. We propose Spatio-Temporal Provenance Guard (STPG), a gateway defense that persists connector provenance, regroups records by source lineage, and authorizes an external routing call only when a trusted directive is witnessed to target that destination and passes a deterministic structural check. Comparisons use evidence adapters on a common kernel, not published defenses. On a Google Cloud testbed with Gemini agents, STPG blocked all 648 attacks and passed 69 of 71 approved transfers. It matched all specified decisions on a policy challenge set and 190/260 on a frozen set from a second model family. Requiring two witness models to agree reduced false authorizations on descriptive notices from 53 to 25 of 150. On a second held-out set generated after all settings were frozen, conformance was 215/260, rising to 229/260 with witness agreement and an extended resolver. Limits remain in reference resolution, descriptive notices, and trusted rewrites, so the results support bounded, destination-level enforcement on synthetic data.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: