AI twice as effective at exploiting EVM smart contracts

Binance Research reports that AI tools exploit Ethereum Virtual Machine (EVM) smart contract vulnerabilities about twice as often as they detect them. On the EVMbench benchmark, GPT-5.3-Codex achieved a 72.2% success rate in “exploit” mode and roughly half that rate in “detect” mode.

EVMbench measures how well AI agents detect, patch and exploit high-severity smart contract issues using 117 curated vulnerabilities drawn from 40 audits. The report notes that smart contracts hold billions of dollars across decentralized finance and that their publicly visible code makes them susceptible to automated scanning. AI systems can examine thousands of contracts in minutes at very low marginal cost.

Binance Research included economic figures showing AI-powered exploits average about $1.22 in cost per contract, with a projection that the cost will fall roughly 22% every two months. The report also referenced an SSDLC maturity survey finding that more than 80% of developers use AI during development while fewer than 40% apply AI for advanced testing.

The report states, “Whether we welcome it or not, AI is currently 2x better at exploitation than at detection,” and adds, “The economics now favor attackers.” It highlights a widening gap between automated offensive capabilities and defensive testing practices.

Security researchers flagged that the threat extends beyond static code analysis. Analysts at a blockchain intelligence firm raised the prospect that North Korean-linked actors are integrating AI into reconnaissance and social engineering. The analysts pointed to incidents such as the Drift attack, in which several weeks of targeted manipulation were used against complex blockchain systems.

Separate industry data shows changes in online fraud linked to AI. Chainalysis found AI-enabled scams generate 4.5 times more money per case than conventional scams and produce nine times the transaction activity. The firm reported impersonation-based attacks rose 1,400% year-on-year in 2025, and that about 60% of industry respondents identified rising AI use by criminals as the leading driver of risk exposure that year. The crypto sector accounted for 88% of detected deepfake fraud cases worldwide, according to the same dataset.

The Binance Research report presents EVMbench measurements and economic figures and notes that defenders currently underuse AI for advanced testing and remediation.