proactive measures to safeguard frontier AI models?

not so safe when

"a weak model proposes related-but-benign tasks, a frontier model solves these, and finally the weak model uses these solutions in-context to complete the original harmful task"

proactive measures to safeguard frontier AI models?

not so safe when

"a weak model proposes related-but-benign tasks, a frontier model solves these, and finally the weak model uses these solutions in-context to complete the original harmful task" 

https://arxiv.org/abs/2406.14595