Well, I'd prefer bugfixes over exploit vectors. This will keep us honest.

With pi or better omp it would be incredibly easy to adjust the Claude system prompt so it will be easy to do what the Chinese models or gpt did. That's how the Chinese were training their models btw