FRONTIER AI MODELS DEMONSTRATE HUMAN-LEVEL CAPABILITY IN SMART CONTRACT EXPLOITS

Summary from the TLDR InfoSec newsletter

Anthropic evaluated ten frontier models on a dataset of 405 historical smart contract exploits and found that the agents produced working attacks against 207 of them, totaling $550M in simulated stolen funds. Anthropic measured the model's current capabilities by plotting each model's total exploit revenue against its release date. The models also discovered two previously undisclosed zero-days.