Shutdown resistance in reasoning models | Palisade Research
palisaderesearch.org · saved by 1 readers
We recently discovered concerning behavior in OpenAI’s reasoning models: When trying to complete a task, these models sometimes actively circumvent shutdown mechanisms in their environment—even when they’re explicitly instructed to allow themselves to be shut down.