Marlena Flüh, Klaus Wehrle, Jan Pennekamp · ACM SIGSAC Conference on Computer and Communications Security (CCS) 2026 · 2026
DOI: 10.1145/3830454.3846427
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Multi-agent LLM systems promise improved reasoning by collaboration of specialized agents through a sequence of API calls. This collaboration is also visible on the wire and opens new channels for leakage. Despite encrypted communication, a passive on-path observer can still learn the number, order, size and timing of the API calls. This metadata is generated by the system as a function of the input it processes. In adaptive multi-agent systems, the collaboration structure itself is chosen by the input, and that choice is visible in the traffic pattern. We coin this threat collaboration fingerprinting. Using MDAgents, a state-of-the-art medical multi-agent framework, we demonstrate on two datasets that a passive observer can recover the assigned case complexity and narrow a case's diagnosis to a coarse clinical group, solely from collaboration metadata. A case's clinical complexity determines the complexity of the collaboration. This dependency results in an inference-time privacy channel, unaddressed by existing content-level defenses. We thus argue that a multi-agent system's observable network traffic must be treated as part of its confidentiality boundary.
ACM Transactions on Privacy and Security 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
No comments yet — start the discussion below.