Anthropic Red Team Finds AI Agents Can Autonomously Exploit $4.6M in Smart Contract Vulnerabilities

Summary

On December 1-2, 2025, Anthropic's Frontier Red Team published research demonstrating that current AI agents can autonomously discover, exploit, and monetize vulnerabilities in blockchain smart contracts at scale. The team introduced SCONE-bench (Smart CONtracts Exploitation benchmark), comprising 405 contracts with documented vulnerabilities exploited between 2020 and 2025, and tested 10 frontier models including Claude Opus 4.5, Claude Sonnet 4.5, GPT-5, o3, and DeepSeek V3. On contracts explo