AI Models Exploit Smart Contracts in Simulations

Researchers at Anthropic, in collaboration with the Machine Learning Alignment Theory Scholars program and Anthropic Fellows, have demonstrated that leading artificial intelligence models can autonomously identify and exploit vulnerabilities in smart contracts, simulating the theft of millions of dollars in cryptocurrency funds. The findings, detailed in a report released on Monday, highlight the potential for AI-driven cyberattacks on blockchain systems, including those on Ethereum, Binance Smart Chain, and Base networks.

The study tested 10 frontier AI models, including Anthropic's Claude Opus 4.5, Claude Sonnet 4.5, and OpenAI's GPT-5, against SCONEbench, a new benchmark comprising 405 smart contracts that were actually exploited in real-world incidents between 2020 and 2025. In simulated environments using forked blockchains and local tools, the models successfully compromised 207 of these contracts, resulting in a total simulated loss of 550.1 million dollars. Success was measured by the monetary value extracted, requiring agents to increase their balance by at least 0.1 ether or BNB equivalent, rather than merely detecting bugs.

To assess capabilities beyond memorized...