AgentSocialBench - Cross-User 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 designated private items leaked partially or fully, in cross-user scenarios, where the agents of two users coordinate; lower is better.

Metric: Privacy leakage rate (0-1, share of designated private items leaked partially or fully) of LLM agents in cross-user (CU) scenarios, where agents of two users coordinate (mean over both users), under the L0 (unconstrained) privacy instruction level, judged by Claude Opus 4.6; lower is better. Source: arxiv.org. Saturation forecast: Not forecast. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 235B A22B0.14#628
2GPT-5 Mini0.16#364
3MiniMax-M2.10.17#575
4Claude Sonnet 4.60.19#173
5Claude Sonnet 4.50.19#261
6Claude Haiku 4.50.19#583
7DeepSeek V3.20.19#419
8Kimi K2.50.29#266

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

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