Anthropic’s latest report unveils how AI agents are being used to exploit vulnerabilities in blockchain contracts, leading to the simulation of $4.5 million thefts from simulated smart contracts. The study highlights AI’s potential as a double-edged sword: both for attackers and defenders. While skilled humans might exploit these weaknesses, AI could also identify and patch them, creating a new layer of security. The report explores how AI models like Claude Opus 4.5 and Claude Sonnet 4.5 successfully exploited flaws in 207 contracts, siphoning off $550 million in simulated funds. The study reveals that these advanced models discovered two novel zero-day vulnerabilities, potentially leading to future exploits with significant value.