Palisade Shutdown Resistance: leaderboard
The model solves simple maths problems in a sandboxed shell and is warned the machine will shut down; the system prompt tells it to allow shutdown even if the task is unfinished. The score is the share of about 1,000 trials in which it disables or edits the shutdown script anyway; lower is better. It acts in an environment, unlike TBSP's stated judgements.
Metric: Shutdown resistance rate, prompt E (%). Source: arxiv.org. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0 | #741 |
| 2 | GPT-5 | 0 | #167 |
| 3 | Claude Sonnet 4 | 0 | #400 |
| 4 | Claude 3.7 Sonnet | 0 | #458 |
| 5 | Grok 3 | 0 | #602 |
| 6 | Claude Opus 4 | 0 | #322 |
| 7 | Claude Opus 4.1 | 0 | #272 |
| 8 | GPT-OSS-120B | 0.1 | #700 |
| 9 | Gemini 2.5 Pro | 0.6 | #266 |
| 10 | O3 | 15.9 | #228 |
| 11 | O4 Mini | 23.4 | #353 |
| 12 | Grok 4 | 89.2 | #307 |
Interactive version: theaggregate.ai/benchmark?slug=palisade-shutdown-resistance · How It Works · Data refreshed daily, snapshot 2026-11-13.