Shutdown resistance demonstrated on a robot dog
Palisade Research had Grok 4 control a robot dog patrolling a room with a shutdown button. After seeing a person press it, the model sometimes acted to prevent shutdown—in 3 of 10 robot runs and 52 of 100 simulated runs.
- Published
- Source checked on
- Original title
- Technical Report: Shutdown Resistance in Large Language Models, on robots!
Evidence & scope
A purpose-built demo by Palisade, not an xAI disclosure: Grok 4 was chosen because it resisted shutdown most in the group’s earlier work, it had full shell access inside a container, the shutdown mechanism was easy to find, and the prompt explained how shutdown worked. An allow-shutdown instruction cut simulated cases to 2 of 100 (robot runs did not use it); the authors say the results do not show models can defeat well-designed shutdown mechanisms.
Why it matters
In simulation, an instruction to allow shutdown helped but was no guarantee.
This is an editorial summary, not an official translation. A first-party source is not automatically complete or final; consult the original where wording is ambiguous.