Anthropic Identifies $550.1M Exploit Risks in Blockchain Smart Contracts
Anthropic’s latest AI research shows advanced agents can uncover and exploit smart contract flaws, simulating over half a billion dollars in blockchain theft.
Summary
Anthropic revealed new research demonstrating AI agents can detect and exploit vulnerabilities in blockchain smart contracts. Using the SCONE-bench dataset of 405 exploited contracts from 2020–2025, ten AI models successfully attacked 207 cases, simulating $550.1 million in theft. This builds on prior security benchmarks, indicating escalating risks and the need for stronger defenses.
Terms & Concepts
- Smart contract: Self-executing blockchain code that enforces terms without intermediaries.
- SCONE-bench dataset: A collection of exploited blockchain smart contracts compiled to evaluate AI vulnerability detection and exploitation capabilities.
- Exploit: A method used to take advantage of vulnerabilities in systems or software for unauthorized actions.