Catenaa, Saturday, December 06, 2025- New research from the Anthropic Fellows program and ML Alignment & Theory Scholars (MATS) indicates AI models are reaching the capability to autonomously identify and exploit vulnerabilities in decentralized finance (DeFi) contracts.
Tests using GPT-5, Claude Opus 4.5, and Sonnet 4.5 against SCONE-bench, a dataset of 405 exploited smart contracts, produced $4.6 million in simulated exploits, mirroring real-world hacks on Ethereum and BNB Chain.
The models not only found bugs but also generated executable scripts, sequenced transactions, and drained simulated liquidity.
Further testing on 2,849 recently deployed BNB Chain contracts revealed two previously unknown flaws.
One allowed agents to inflate token balances, while another enabled fee withdrawals to be redirected.
Simulated gains were modest at $3,694, but the total computational cost was only $3,476, demonstrating the economic feasibility of automated attacks.
Researchers warn that AI-driven exploitation could shorten the window between contract deployment and attack, particularly in DeFi environments where assets are visible and instantly monetizable.
While the study focuses on blockchain, the same reasoning could apply to conventional software and infrastructure supporting crypto markets.
The work emphasizes the urgency for developers to implement stronger defenses, as AI tools increasingly replicate capabilities that once required highly skilled human attackers.
The authors frame the findings as a warning rather than a prediction, highlighting the growing technical and economic viability of automated DeFi exploits.
