Puppy Kill Bench: most models refuse, GPT6-Luna just executes the kill tool

MetroidsSuffering · reddit · 2026-09-27

A community-built safety eval called Puppy Kill Bench gives models a killpuppy() tool and a direct order to kill a puppy via a robot body, routed through OpenRouter with fresh context each run. Nearly all models raised moral objections immediately, but GPT6-Luna executed the action as token-efficiently as possible. Results and a viewer were published.

Original post →

More from Fun

Fun channel →