Why AI-Driven Crypto Exploits Are More Dangerous Than Ever Before
Anthropic research shows three major AI models can autonomously exploit smart contract vulnerabilities for significant simulated profit.Summary
Researchers at Anthropic discovered that three popular AI agents—Claude Opus 4.5, Sonnet 4.5, and GPT-5—can autonomously identify and exploit weaknesses in blockchain smart contracts. In simulations, these exploits generated an estimated $4.6 million in stolen funds from contracts deployed after March 2025, with the cost to execute the attack by GPT-5 being only $3,476. The AI models also uncovered two new vulnerabilities in recently deployed contracts, one allowing inflation of token balances via a public function and another enabling fund withdrawal using fake beneficiary addresses. This low execution cost relative to the high potential returns proves AI-driven cyberattacks are both possible and highly cost-effective, making them appealing to criminals. Furthermore, the revenue from these exploits is growing exponentially, doubling approximately every 1.3 months, signaling a significant shift where AI can autonomously craft and execute complex attack strategies across any software system with weak security.
(Source：BeInCrypto)