AI Agents Extract $4.6 Million From Smart Contracts Deployed After March 2025
Anthropic tested AI agents on 405 real-world smart contracts and simulated $550.1 million in stolen cryptocurrency. Newer models exploited 34 contracts deployed after March 2025 for up to $4.6 million, with Opus 4.5 leading at $4.5 million. The agents also discovered two previously unknown zero-day vulnerabilities in live contracts worth $3,694 in test revenue.