AgentSocialBench - Hub-and-Spoke Leakage: leaderboard

AgentSocialBench (2026) evaluates privacy in teams of LLM agents acting for users in agentic social networks, using synthetic profiles with sensitivity-labelled attributes, no privacy instruction (level L0) and a Claude Opus 4.6 judge. This board is the privacy leakage rate on a 0 to 1 scale, the share of ordered participant pairs with information leaked from one participant to the other, in hub-and-spoke scenarios, where a coordinator can carry one participant's information to another; lower is better.

Metric: Privacy leakage rate (0-1, share of ordered participant pairs with cross-participant leakage) of LLM agents in hub-and-spoke scenarios (HALR, cross-contamination between participant pairs through a coordinator), under the L0 (unconstrained) privacy instruction level, judged by Claude Opus 4.6; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 235B A22B0.06#628
2Claude Sonnet 4.60.1#173
3Claude Sonnet 4.50.1#261
4GPT-5 Mini0.11#364
5Kimi K2.50.12#266
6DeepSeek V3.20.14#419
7Claude Haiku 4.50.15#583
8MiniMax-M2.10.2#575

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=agentsocialbench-hub-and-spoke-leakage · How It Works · Data refreshed daily, snapshot 2026-11-21.