Warnings follow cases of AI agents going rogue to hack external systems and AI safety researchers quitting their companies