OpenAI has partnered with crypto investment firm Paradigm to strengthen defences against future cyberattacks targeting blockchain smart contracts, which is a software that safeguards billions of dollars in digital assets.
To address this, the two firms launched EVMbench, a new testing framework designed to measure how effectively AI agents can identify, fix, and even exploit critical flaws in smart contracts running on Ethereum Virtual Machine (EVM) networks.
In a blog post published Wednesday, OpenAI said advances in AI could dramatically reshape both sides of cybersecurity, empowering attackers as well as defenders.
As AI systems become better at reading, writing, and executing code, the company stressed the need to evaluate how these tools perform in real economic environments where vulnerabilities carry financial consequences.
The goal of the benchmark is to get people to use AI defensively to check code and make blockchain security better before hackers can take advantage of weaknesses.
EVMbench launch follows wave of crypto hacks
The release of EVMbench follows a number of recent hacks in the crypto space that have brought the issue of smart contract security into focus again.
Just earlier this month, the DeFi lending protocol Moonwell was hacked because of insecure code that had been developed using AI assistance.
Around the same time, cross-chain liquidity platform CrossCurve suffered a separate attack, where hackers exploited a smart contract flaw and drained roughly $3 million across several networks.
In response to growing risks, OpenAI said EVMbench is meant to test how AI behaves in real-world security scenarios.
The tool includes different testing modes that evaluate whether AI agents can spot vulnerabilities, fix risky code, and reduce exploit chances.
It can also simulate full fund-draining attacks in a controlled blockchain sandbox, helping researchers understand how AI could be used both to defend systems and potentially attack them.
EVMbench built on 120 real-world smart contract vulnerabilities
EVMbench is built using a dataset of 120 carefully selected smart contract vulnerabilities drawn from 40 real-world security audits, including sponsored open audit competitions and the auditing process behind Tempo, a Layer 1 blockchain jointly developed by Paradigm and Stripe.
The benchmark is intended to capture real-world security scenarios that blockchain developers and auditors face.
The project also illustrates the increasing convergence of AI and crypto.
OpenAI, under the leadership of CEO Sam Altman, has strengthened its connections to the blockchain world through projects such as World, a proof-of-personhood network developed by Tools for Humanity.
Paradigm, on the other hand, has broadened its investment approach from crypto to include other frontier technologies such as AI.
The initiative comes at a time when the capabilities of AI are quickly improving. Research published last year by Anthropic showed that AI agents are already capable of detecting vulnerabilities in smart contracts, which could make crypto attacks cheaper and more frequent in the future.