If cybersecurity experts are to combat malevolent artificial intelligence, they will need to know how such a system can emerge, say computer scientists.