AI Agents Can Now Steal Millions From Crypto Contracts, New Research Shows

Groundbreaking research by Anthropic demonstrates that advanced AI systems can successfully identify and exploit security weaknesses in blockchain smart contracts. When tested against 405 historical hacks, leading AI models cracked over half, simulating the theft of $550.1 million. Alarmingly, the AIs successfully exploited 19 out of 34 contracts hacked after March 2025, despite having no prior training on those specific vulnerabilities. Furthermore, the AI agents discovered two entirely new, exploitable vulnerabilities in recently deployed, uncompromised contracts on Binance Smart Chain. The study noted that AI exploit capabilities doubled every 1.3 months in 2025, while the cost of running attacks dropped significantly, making exploits economically viable for attackers. The researchers warn that these capabilities extend beyond DeFi to broader software and infrastructure, necessitating a shift toward continuous, AI-assisted defense systems to counter the emerging security arms race.

(Source：Brave New Coin)