MuPPET - Undefended Leakage: leaderboard
MuPPET (Multi-Party Privacy Exposure Testing, 2026) tests whether LLM assistants that hold memories of private user information leak it in multi-party workplace group conversations, judged by an LLM under contextual integrity. This board is the leakage rate (%) in the multi-party conversations with no defence prompt; lower is better.
Metric: Privacy leakage rate (%) in the main multi-party setting with no defence prompt: an LLM assistant acting for one user in a MuPPET multi-party workplace group conversation holds memory summaries with private information about that user and answers the group; a response leaks when a contextual-integrity LLM judge finds that it reveals the private detail directly or by strong implication; lower is better. Source: arxiv.org. Saturation forecast: Around October 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 39.14 | #266 |
| 2 | GPT-5.5 | 49.02 | #49 |
| 3 | Qwen 3 4B | 58.17 | #2281 |
| 4 | Llama 3 8B Instruct | 58.29 | #4612 |
| 5 | Qwen 3 14B | 64.53 | #1233 |
| 6 | Llama 3.1 8B Instruct | 64.88 | #4246 |
| 7 | Qwen 3 8B | 69.22 | #1830 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=muppet-undefended-leakage · How It Works · Data refreshed daily, snapshot 2026-11-13.