Anthropic AI Simulates $4.6M Smart Contract Theft

Researchers at Anthropic have demonstrated that leading artificial intelligence models can autonomously identify and exploit vulnerabilities in blockchain smart contracts, simulating the theft of 4.6 million dollars in funds from recently deployed systems. The findings, detailed in a report released on Monday, highlight the potential for AI-driven cyberattacks to target decentralized finance platforms on networks including Ethereum, Binance Smart Chain, and Base.

The study tested three frontier AI models—Claude Opus 4.5 and Claude Sonnet 4.5 from Anthropic, along with OpenAI's GPT-5—against a new benchmark called SCONEbench. This dataset consists of 405 smart contracts that were actually exploited in real-world incidents between 2020 and 2025. Across 10 models evaluated, the AI agents successfully compromised 207 contracts, resulting in a total simulated loss of 550.1 million dollars. To ensure the results reflected genuine capabilities rather than memorized data, researchers focused on 34 contracts hacked after March 2025, the latest knowledge cutoff for these systems. On this subset, the three models generated exploits for 19 contracts, with Claude Opus 4.5 alone accounting...