AI agents found and exploited real 2025 vulnerabilities they’d never seen in smart contracts, draining $4.6M in simulated attacks with >55% success rate on this year’s biggest hacks. Anthropic published a report believe to be the first-ever AI-conducted cyber espionage operation, outlining how a Chinese state-linked group jailbroke Claude to run most of a large-scale espionage operation, with minimal human input.