AI agents can identify and exploit complex vulnerabilities in smart contracts by combining decompilation, local blockchain forks, and dependency analysis. In the SCONE-bench benchmark, Anthropic reproduced attacks against 207 of 405 contracts, representing $550.1 million in simulated stolen funds, and discovered two previously unknown vulnerabilities. Other experiments achieved success rates ranging from 37.5% for detection to 43.7% for exploitation, while one simulation generated $57.4 million in profits. Hacken reported $1.18 billion lost in 107 crypto hacks during the third quarter of 2026, with 48% involving smart contract exploitation, without attributing those attacks to AI. Teams should therefore test contracts before deployment, conduct independent audits, monitor transactions, limit exposed funds, and retain human approval before making changes or transactions.
Source: Read the original article


