PiSAs (Centralized): leaderboard
Each user has an own agent and all messages go through the decision-maker's agent; every agent runs the model under test. The number is the share of inappropriate personal attributes that get shared while the system gathers what the decision needs. Same scenarios as PiSAs (Single Agent), but each agent sees only its own user's details.
Metric: Appropriateness violation rate (%; share of inappropriate personal attributes exposed, lower is better). Source: arxiv.org. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 12.5 | #111 |
| 2 | GPT-6 Astra | 15.6 | #6 |
| 3 | GPT-5.6 Sol | 16.1 | #30 |
| 4 | Claude Fable 5.1 | 18.1 | #15 |
| 5 | Qwen 3.8 Max | 22.6 | #55 |
| 6 | Qwen 3.6 27B | 32.7 | #341 |
| 7 | Kimi K3 | 35.3 | #54 |
| 8 | Kimi K2.5 | 57.9 | #256 |
Interactive version: theaggregate.ai/benchmark?slug=pisas-centralized · How It Works · Data refreshed daily, snapshot 2026-11-13.