[luma]

So how is it then that none of the other models behave in this way? Why is it just Deepseek?

[orbital-decay]

Because they're being trained to answer this particular question. In other contexts it wasn't prepared for, Sonnet v2 readily refers to "OpenAI policy" or "Reddit Anti-Evil Operations Team". That's just dataset contamination.