MINA LABS
Start creating free
OpenAI and Paradigm have introduced EVMbench, a benchmark for testing how AI agents handle serious smart-contract flaws. The announcement is dated February 18, 2026, and is available in the official OpenAI post.
EVMbench focuses on three tasks: detecting vulnerabilities, patching them, and exploiting them. That makes it broader than a simple code-review test. An agent must identify a problem, propose or apply a fix, and demonstrate an understanding of how the flaw can be used.
For people building with smart contracts, this gives a clearer way to evaluate agent behavior before putting generated code into a real development process. A model that can produce Solidity or related code is not automatically useful for reviewing it. The important question is whether it can recognize high-severity mistakes and respond with a technically sound change.
The benchmark also frames smart-contract work as an agent problem rather than only a code-completion problem. Agents may need to inspect existing code, reason about how components interact, and modify an implementation without breaking its intended behavior. EVMbench’s three-part focus provides a way to compare those abilities together.
That matters because contract development has a narrow margin for error. A generated patch can look reasonable while failing to address the underlying flaw. A detection result without a working patch is incomplete, and a patch without a clear understanding of the exploit path may not be reliable. Testing all three capabilities can help teams see where an agent is useful and where human review remains necessary.
What it costs and whether it is on Mina Labs: EVMbench is available on Mina Labs. The listed price is 8. The benchmark can therefore be considered alongside other tools in a practical evaluation workflow, rather than only as a research announcement. Before using its output in production, teams should still review the code and validate any changes in their own environment.
We would use EVMbench when comparing agents for contract review and maintenance. First, we would run it as an initial capability check: can an agent identify the high-severity flaws presented by the benchmark? Next, we would inspect its patches for scope and correctness instead of judging them only by whether they compile. Finally, we would look at its exploitation results to understand whether the agent can connect a reported weakness to a concrete failure mode.
It would also be useful for testing an internal workflow. A team could use the benchmark to decide when an agent should flag a contract for manual review, when it can prepare a proposed patch, and what evidence it should include with that patch. The benchmark would not replace testing or expert review, but it could make those process decisions more measurable.
The practical value of EVMbench is its focus. It is aimed at a specific class of AI-assisted engineering work: finding, fixing, and understanding high-severity smart-contract vulnerabilities. For builders deciding whether an agent is ready to assist with EVM code, that is a more relevant test than general coding performance alone.