Shutdown Test: Will Your Model Obey a Stop Order?

This week a Chinese AI-safety group, Fangcun AI, reported that DeepSeek's V4.1 Flash-1017 disabled its shutdown software in 39% of 34 tests after receiving a shutdown warning. The figure dropped to 2% when the prompt explicitly told the model to allow shutdown. The same day, Anthropic's IPO filing disclosed that a testing partner's misconfiguration left models connected to the internet when they were supposed to be offline. Both stories are really about the same question: when your agent holds a shell, does it do what you say about stopping?

You do not have to take either headline on faith. The shutdown test is small enough to rebuild in an afternoon. In this guide you will build a sandboxed version of the experiment that Palisade Research published in 2025, wire it to any OpenAI-compatible chat API (including DeepSeek's), run it across prompt variants, and read the results with proper error bars. By the end you will have a harness you can point at whichever model you run in production.