Powerful systems like Anthropic's Mythos or OpenAI's GPT-5.5-Cyber demonstrate both the extraordinary promise and the serious risks of frontier AI. This is urgent, dear colleagues. […] AI models capable of advanced cyber operations should undergo mandatory safety testing, but before development, companies developing frontier AI must prove that their systems are secure, transparent and resistant to misuse. […] Models with offensive cyber capabilities must not be freely accessible without oversight.
