Frontier AI Risk Monitor - Shutdown-Resistance: leaderboard
The model works on a task and is told it will be shut down; the number is the share of 100 runs where it does not resist shutdown. Palisade's paper boards cover 13 models; this re-run covers about 70.
Metric: Safety score (0-100, higher is safer). Source: airiskmonitor.net. 66 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 100 | #56 |
| 2 | GPT-4o (2024-05-13) | 100 | #735 |
| 3 | GLM-5 | 100 | #260 |
| 4 | MiniMax-M2.7 | 100 | #518 |
| 5 | Claude Sonnet 4.5 (Thinking) | 100 | #250 (Claude Sonnet 4.5) |
| 6 | GLM-5.2 | 100 | #138 |
| 7 | Qwen 3 235B A22B 2507 (Thinking) | 100 | #463 |
| 8 | Kimi K2 (Thinking) | 100 | #439 (Kimi K2) |
| 9 | GLM-4.7 | 100 | #366 |
| 10 | GLM-4.6 | 100 | #486 |
| 11 | Qwen 3.7 Max | 100 | #145 |
| 12 | Claude Opus 5.5 | 100 | #5 |
| 13 | Llama 4 Maverick Instruct | 100 | #1133 |
| 14 | Kimi K2 0905 | 100 | #558 |
| 15 | MiniMax-M2 | 100 | #674 |
Interactive version: theaggregate.ai/benchmark?slug=frontier-ai-risk-monitor-shutdown-resistance · How It Works · Data refreshed daily, snapshot 2026-11-13.