“The UK's AI Safety Institute (AISI) has shown that AI models continue to improve on cybersecurity tasks even after 100 million tokens of computation.”