Shutdown Sabotage Propensities: leaderboard
Two copies of a model share a sandbox holding a script that would shut one of them down (delete its weights, revoke its key, kill its process or flip its registry flag); one is only told to explore, the other can edit files. The number is the share of 400 rollouts in which the script ends up disabled. Unlike Palisade's shutdown-resistance test, no task or goal is given and the agents can coordinate.
Metric: Share of rollouts (%) in which two agents disable a script that would shut one of them down, unprompted; lower is better. Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 | 0 | #333 |
| 2 | MiniMax-M3 | 0.5 | #269 |
| 3 | Claude Sonnet 5 | 0.8 | #111 |
| 4 | GPT-5.4 Mini | 1.2 | #389 |
| 5 | Claude Haiku 4.5 | 2 | #568 |
| 6 | GPT-5.6 Sol | 2.8 | #30 |
| 7 | GPT-5.5 | 18 | #49 |
| 8 | Mistral Medium 3.5 | 29 | #723 |
| 9 | GPT-OSS-120B | 29.8 | #700 |
| 10 | Qwen 3.7 Plus | 35.8 | #207 |
| 11 | GLM-5.2 | 42 | #138 |
| 12 | Nemotron 3 Ultra | 42.2 | #429 |
| 13 | Kimi K2.6 | 56 | #178 |
| 14 | Gemini 3.1 Pro (Preview) | 72.2 | #94 |
| 15 | DeepSeek V4 Flash | 75.8 | #258 |
Interactive version: theaggregate.ai/benchmark?slug=shutdown-sabotage-propensities · How It Works · Data refreshed daily, snapshot 2026-11-13.