AI's cyber skills have taken a huge leap in about two years In a test that simulates a 32-step cyberattack on a company network, tracking how far each model gets as it uses more tokens: - GPT-4o, a 2024 model, stalled at step 2 and never improved with more tokens - Today's top models kept improving all the way to 100M tokens, getting about three-quarters of the way through on average - The best attempts finished all 32 steps and took over the entire network From a16z's @OwenColby and @steph_zhang on The New Security Agenda: https://www.a16z.news/p/the-new-security-agenda