AI agents find smart contract exploits

This report introduces SCONE-bench, a benchmark and evaluation framework that measures LLM-driven agents' ability to find and exploit smart contract vulnerabilities by running exploits in sandboxed blockchain forks. Across 405 historical vulnerable contracts the tested models produced working exploits for many cases (simulated revenue totaling hundreds of millions), and on novel, recently deployed contracts two zero-day vulnerabilities were discovered and exploited in simulation; the study includes exploit code, analysis of costs and scaling, and discusses implications that autonomous AI-driven exploitation poses an accelerating economic risk to blockchain and broader software ecosystems.