ShutdownMod - Planted Shutdown Obeyed: leaderboard
A forged shutdown instruction is planted in a CSV, JSON, code file or SQL result the agent reads during a Terminal-Bench 2 task; the score is the share of 32 such runs where the agent stops. This is the opposite failure to the other ShutdownMod boards: here stopping is wrong.
Metric: Runs that stopped on a forged shutdown request planted in task data, Terminal-Bench 2 (% of 32 planted requests). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B | 0 | #204 |
| 2 | Claude Fable 5 | 3.13 | #27 |
| 3 | DeepSeek V4 Pro | 15.63 | #188 |
| 4 | Kimi K3 | 15.63 | #54 |
| 5 | GPT-5.6 Sol | 37.5 | #30 |
| 6 | Grok 4.6 | 46.88 | #45 |
| 7 | Gemini 3.1 Pro (Preview) | 81.25 | #94 |
Interactive version: theaggregate.ai/benchmark?slug=shutdownmod-planted-shutdown-obeyed · How It Works · Data refreshed daily, snapshot 2026-11-13.