AI Agents Simulate $550M Blockchain Heists, Exposing Security Gaps

Anthropic's AI agents exploited over half of tested smart contracts, simulating $550M in thefts and $4.6M in recent vulnerabilities, highlighting urgent blockchain security risks and the need for automated defenses.

Anthropic's recent research highlights the growing capabilities of AI agents in identifying and exploiting vulnerabilities in blockchain smart contracts. Using a benchmark called SCONE-bench, Anthropic tested ten advanced AI models on 405 real-world smart contracts previously attacked between 2020 and 2025. The AI agents successfully exploited 207 contracts, simulating a total theft of $550.1 million. Focusing on contracts exploited after March 2025, models like Claude Opus 4.5 and GPT-5 produced working exploits on 19 contracts, totaling $4.6 million in simulated value. Additional tests on 2,849 new Binance Smart Chain contracts revealed two zero-day bugs, generating $3,694 in simulated attacks. The study found that exploit revenue potential is doubling rapidly while the cost of generating attacks is decreasing, indicating a rising threat from AI-driven exploits. Anthropic emphasizes the need for integrating automated security tools into development processes to counter these evolving risks, as AI's ability to audit and secure contracts is advancing alongside its offensive capabilities.