Aidan Ewart comments on Red-teaming language models via activation engineering