AI agents are getting good enough at finding attack vectors in smart contracts that they can already be weaponized by bad actors, according to new research published by the Anthropic Fellows program.