Do you think the training process will retain these artifacts? I doubt it. If they were simply stealing the content - sure it would make sense - but I suspect they’re feeding it into training data and RL might distill these out.
https://www.anthropic.com/research/small-samples-poison?from...