Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over gDP forecasting
Import AI Jack Clark
Researchers at Lyptus found that AI systems' ability to perform cyberattacks improves predictably with each model generation, with capability doubling time accelerating from 9.8 months (since 2019) to 5.7 months (since 2024). The most advanced current models achieve 50% success on hacking tasks that require human experts 3.2 hours to complete. Open-weight models are catching up to closed-source frontier models on a timeline of roughly 5.7 months, raising concerns about diffusion of offensive cyber capabilities.
Why it matters
How much could AI revolutionize the economy?
Related stories
Import AI 455: AI systems are about to start building themselves.
Import AI · 3 months ago ·
24
Chinese AI models narrow cyber gap with US rivals
CSET Georgetown · 1 month ago ·
8
Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems
Import AI · 3 months ago ·
30
Chinese AI models narrow cyber gap with US rivals
CSET Georgetown · 1 month ago ·
8
Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems
Import AI · 3 months ago ·
30