Catenaa, Saturday, February 28, 2026- OpenAI and crypto investor Paradigm have partnered to develop an AI agent tool aimed at strengthening smart contract security, the companies announced Wednesday.
The initiative introduces EVMbench, a benchmark designed to evaluate AI agents’ ability to detect, patch, and exploit vulnerabilities in smart contracts.
The collaboration seeks to address growing concerns over DeFi exploits. Earlier this month, Moonwell suffered a $1.78 million loss after a smart contract misconfiguration, and CrossCurve recently lost roughly $3 million due to another contract vulnerability.
EVMbench will simulate real-world attack scenarios in a sandboxed blockchain environment, allowing AI agents to identify high-severity flaws and test mitigation strategies without endangering live networks.
EVMbench draws from 120 curated vulnerabilities collected from 40 audits, including sponsored competitions and the Tempo blockchain security process co-developed by Paradigm and Stripe.
OpenAI said the tool features multiple capability modes, including vulnerability detection, contract modification, and executing end-to-end fund-draining attacks to assess defensive performance.
OpenAI CEO Sam Altman emphasized the need for AI systems to be used defensively to audit and strengthen deployed contracts. Paradigm has increasingly focused on frontier technologies, including AI, to anticipate emerging risks in the crypto ecosystem.
The launch comes amid rising concern that AI-assisted coding could accelerate both attacks and defenses in decentralized finance.
Competing AI research firm Anthropic noted that agents are already capable of identifying vulnerabilities in contracts, potentially lowering the cost of exploits.
EVMbench aims to provide measurable benchmarks and testing environments to improve resilience before assets are deployed onchain.
