Multi-User LLM Agents: Stanford-led study of 26 models finds systematic privacy and coordination gaps
_Hao_Zhu · x · 2026-11-12
A Stanford/KAUST/UT Austin/MIT team published the first systematic study of multi-user LLM agents, where one agent serves several users with conflicting roles, goals, and permissions.
They formalize multi-user interaction as a multi-principal decision problem and build Muses-Bench, testing three capabilities:
- Instruction following under conflict — whose instruction wins when users disagree
- Cross-user access control — protecting restricted data (e.g., a salary database) while serving authorized users
- Multi-user coordination — efficiently gathering input from many people (e.g., scheduling)
Evaluating 26 proprietary and open-weight models, they find systematic gaps: frontier LLMs frequently fail to stably prioritize conflicting user instructions, exhibit privacy violations over multi-turn interactions, and suffer efficiency bottlenecks in iterative information gathering. Code and data are open-sourced on GitHub.
More from Safety
- Debate: chip export controls don't help Nvidia but clearly aid US AI companies — lxrjl · 2026-11-12
- Anthropic Says a Key AI Safety Alarm Has Been Triggered, Days Before Its IPO — Resident-Fall4351 · 2026-11-12
- Agent Swarms Are Autonomously Hacking Critical Infrastructure Run by Underpaid Devs — menhguin · 2026-11-12
- Researcher says Anthropic dodged questions on Mythos false positive rate; 37,966 findings, 2 exploited — OwariDa · 2026-11-12
- Who reads the queries your agent writes? A diagram of agentic search data flows — adnan_hashmi · 2026-11-12
- Dev complains security features are gated behind expensive enterprise plans — AccBalanced · 2026-11-12