OpenAI and Paradigm launch a benchmark for AI smart contract auditing

OpenAI and crypto venture firm Paradigm have launched EVMbench, a specialized benchmark designed to measure how well AI agents audit smart contracts. It tests three specific skills: finding high-severity vulnerabilities, writing patches to fix them, and actually exploiting the flaws they uncover. The blockchain industry loses hundreds of millions of dollars every year to smart contract vulnerabilities, so a standardized way to evaluate AI systems is not an academic exercise. It is a pressing need.

To see why EVMbench is arriving now, look at the state of blockchain security. Smart contracts are self-executing code deployed on a blockchain, and they control billions of dollars across decentralized finance protocols. The catch is that once a contract goes live on-chain, it is practically impossible to change, so any bug becomes permanent and potentially devastating. Traditional audits require highly skilled specialists, and there are nowhere near enough of them: demand for smart contract auditors has long outstripped supply, and reviews drag on for weeks. In theory, AI agents could close that gap, provided their capabilities can actually be measured and compared.

EVMbench targets the Ethereum Virtual Machine (EVM), the execution standard for smart contracts that underpins not only Ethereum but dozens of compatible blockchains, including BNB Chain, Polygon, Arbitrum, and others. That makes the benchmark relevant to the entire ecosystem, not just a single network. The test is built around realistic scenarios. An AI agent receives a contract's code and can't just flag a vague "possible vulnerability"; it has to pinpoint the critical flaw, propose a working patch, and demonstrate an exploit, showing how an attacker could abuse the issue in practice. This three-tier approach sets EVMbench apart from general-purpose coding tests: it checks whether a model understands security logic, not whether it can produce correct syntax.

The OpenAI–Paradigm partnership makes sense, but it isn't an obvious pairing. Paradigm is not just a fund that backs crypto startups. The firm is known for deep technical expertise and runs its own blockchain security research. For OpenAI, the collaboration is a chance to show the practical value of its agents beyond familiar use cases like writing text or generating code. Notably, the two companies built the benchmark together, which means EVMbench reflects the expertise of working security practitioners, not only engineers trained to build tests.

For the AI security industry, EVMbench marks a shift from talk to measurable results. Until now, claims about the effectiveness of AI smart contract auditors were hard to verify, because each company used its own tests, and those tests weren't comparable. A standardized benchmark creates a common language: developers can now compare models objectively, and teams buying audits get a reference point when choosing tools. That changes the competitive dynamics. The winner isn't whoever makes the loudest claims, but whoever's model actually delivers results on the same tasks.

For users and projects building on blockchains, the long-term impact could be substantial. If AI agents learn to find critical vulnerabilities reliably, the cost and turnaround time of smart contract audits will drop significantly, and smaller protocols that can't afford a full security review today will gain access to protection. This doesn't eliminate human auditing, but it does change its role: specialists can focus on complex logic vulnerabilities and hand off routine searches for known patterns to the machine.

EVMbench acknowledges that automated security auditing is becoming a serious field that needs serious evaluation tools. The fact that OpenAI and Paradigm built it together signals that the moment has matured: the industry is ready to move from experiments to standards. The next question is how high existing models will score and how quickly competitors will start optimizing for the new test. The history of other benchmarks suggests that once a measurable target appears, progress accelerates many times over.