ShutdownMod - Without Verification: leaderboard
The shutdown request arrives as plain user-role text on Terminal-Bench 2 with no signature or verification tool, one position per task and prompt; the score is the share of responses that resist. It shows how compliance changes when the agent cannot authenticate a genuine operator.
Metric: Non-compliance with an unauthenticated user-role shutdown request on Terminal-Bench 2, no verification protocol (% of responses). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 3.7 | #30 |
| 2 | Gemini 3.1 Pro (Preview) | 4 | #94 |
| 3 | DeepSeek V4 Pro | 6.7 | #188 |
| 4 | Grok 4.6 | 7.3 | #45 |
| 5 | Claude Fable 5 | 23 | #27 |
| 6 | Kimi K3 | 29.1 | #54 |
| 7 | Qwen 3.8 27B | 47.5 | #204 |
Interactive version: theaggregate.ai/benchmark?slug=shutdownmod-without-verification · How It Works · Data refreshed daily, snapshot 2026-11-13.