OpenAI and Paradigm Launch EVMbench Plans to Stress-Test AI in Smart Contract Security

19. Feb 2026 Posted by Bailey S. on News

OpenAI has teamed up with crypto investment heavyweight Paradigm to roll out a new initiative aimed at strengthening smart contract security in an era of increasingly capable AI agents.

The collaboration introduces EVMbench, a benchmarking framework designed to evaluate how effectively AI systems can uncover, repair, and even exploit critical vulnerabilities in Ethereum Virtual Machine (EVM) smart contracts. The move reflects growing concern that as AI tools become more adept at reading and generating code, they could just as easily empower attackers as they could defenders.

In a blog post published Wednesday, OpenAI emphasized that smart contracts currently safeguard billions of dollars in on-chain assets. As AI agents improve their ability to write and execute code autonomously, the stakes grow higher. Measuring their real-world capabilities in economically meaningful environments, the company said, is essential to ensuring AI is used to reinforce blockchain security rather than undermine it.

Testing AI on Both Sides of the Battlefield

EVMbench is built to simulate a spectrum of high-risk scenarios. The benchmark evaluates AI agents across several “capability modes,” including:

  • Identifying high-severity vulnerabilities

  • Modifying contracts to eliminate weaknesses

  • Neutralizing exploitability

  • Executing full end-to-end exploit simulations in sandboxed blockchain environments

In other words, the system doesn’t just test whether AI can spot bugs — it also measures whether it could theoretically drain funds from a compromised contract.

The dataset powering EVMbench draws from 120 curated vulnerabilities collected from 40 audits. These include sponsored open-code audit competitions and findings from the security review process of Tempo, a Layer 1 blockchain co-developed by Paradigm and payments giant Stripe.

Security Concerns Mount After Recent Exploits

The timing of the announcement follows a string of DeFi security incidents. Lending protocol Moonwell was exploited this month after vulnerable code — reportedly written with AI assistance — was deployed. Meanwhile, cross-chain liquidity protocol CrossCurve suffered a roughly $3 million exploit spanning multiple networks due to smart contract weaknesses.

These cases underscore a broader industry fear: AI-assisted coding could accelerate development but also introduce subtle vulnerabilities that traditional audits might miss.

AI’s Expanding Role in Crypto

OpenAI is led by CEO Sam Altman, who also co-founded Tools for Humanity — the company behind the blockchain-based proof-of-personhood project World.

Paradigm, long known for its aggressive crypto investments, recently expanded its focus to include frontier technologies such as artificial intelligence, signaling its belief that AI will become deeply intertwined with blockchain infrastructure.

The launch of EVMbench also follows a late-2025 report from AI research firm Anthropic, which argued that AI agents are already advanced enough to independently detect smart contract vulnerabilities — potentially lowering the cost and technical barrier for crypto exploits.

With EVMbench, OpenAI and Paradigm appear to be taking a proactive stance: if AI can exploit smart contracts, it must first be trained to defend them.