Anthropic recently revealed its SCONE-bench, a benchmark designed to evaluate AI models’ effectiveness in identifying vulnerabilities in blockchain smart contracts. By leveraging its Claude AI models, Anthropic uncovered an attack potentially worth $4.6 million, but chose not to exploit it, emphasizing ethical considerations in AI usage. The SCONE-bench dataset includes 405 smart contracts from Ethereum and other blockchains, reflecting the growing sophistication of AI in finding security flaws. Researchers noted a significant increase in exploit revenue, doubling every 1.3 months. In controlled tests, Anthropic’s AI models, particularly GPT-5, effectively identified vulnerabilities in new contracts, highlighting both the feasibility of autonomous exploitation and the urgent need for proactive AI defenses. As the cost of identifying vulnerable contracts drops, the threat landscape intensifies, necessitating advanced security measures. This underscores AI’s dual role in both creating and defending against cybersecurity risks in the evolving blockchain ecosystem.
Source link
