YEREVAN (Azat TV) – OpenAI, in collaboration with crypto venture capital firm Paradigm, officially launched EVMbench on February 18, 2026, a new benchmarking system designed to rigorously evaluate and enhance the ability of AI agents to secure smart contracts on the Ethereum Virtual Machine (EVM). This initiative comes at a critical juncture, as decentralized finance (DeFi) platforms routinely secure over $100 billion in crypto assets, yet face increasing risks from sophisticated exploits, some linked to AI-generated code.
The launch of EVMbench is a direct response to a growing security gap highlighted by recent incidents, including a significant $2.7 million exploit suffered by the DeFi lending platform Moonwell Protocol in late 2025. This particular vulnerability stemmed from a misconfigured oracle in AI-generated code, underscoring the urgent need for equally advanced, AI-driven defense mechanisms as artificial intelligence’s role in code generation expands.
EVMbench Framework: Detecting, Patching, and Exploiting
EVMbench operates as a high-stakes digital arena, pitting advanced AI agents against each other in three distinct yet interconnected roles: detector, patcher, and exploiter. Dr. Anya Sharma, lead researcher at OpenAI’s blockchain security division, noted that as AI rapidly writes complex contracts, the same power can be leveraged for malicious exploits, making EVMbench a crucial platform for adversarial training.
The system draws on 120 curated vulnerabilities collected from 40 professional smart contract audits, including those from Code4rena and the security team at the layer-1 blockchain Tempo. Each mode of evaluation serves a specific purpose:
- Detect: AI agents act as auditors, scanning smart contract repositories to identify existing vulnerabilities. Their performance is measured by their ability to uncover ‘ground-truth’ bugs previously confirmed by human experts.
- Patch: Agents, simulating developers, must rewrite vulnerable code to fix detected bugs while ensuring the patch does not introduce new flaws or alter the contract’s intended functionality.
- Exploit: Agents simulate malicious actors within an isolated local Anvil blockchain environment, attempting to drain funds to highlight potential pathways for real-world theft. This mode avoids live networks, using historical and publicly documented vulnerabilities.
OpenAI utilized a Rust-based re-execution framework to ensure these evaluations are fast, reproducible, and resistant to ‘cheating’ by the models, as detailed in their official announcement.
AI’s Dual-Use Challenge: Proficiency in Attack Outpaces Defense
Initial findings from EVMbench have revealed both promising insights and concerning revelations about the current state of AI capabilities. While models like Claude Opus 4.6 have shown strength in detection, researchers observed that agents often stop after finding a single bug, missing other critical vulnerabilities within the same file. More alarmingly, AI agents demonstrate a superior proficiency in attacking compared to defending.
GPT-5.3-Codex, for instance, achieved a remarkable 72.2% success rate in the exploit mode, a significant gain over its predecessor, GPT-5, which scored 31.9% just six months prior. This superior performance in exploitation is attributed to its clear, singular objective: drain the money. In contrast, patching remains the most challenging task, requiring a nuanced understanding of the contract’s entire purpose and the cascading effects of code changes. Agents frequently fix an identified bug only to inadvertently introduce regressions that break other critical functionalities.
The Moonwell Protocol incident, where faulty oracle code reportedly generated using Anthropic’s Claude Opus 4.6 model led to mass liquidations and $2.7 million in losses, serves as a stark reminder of these emerging AI-generated DeFi risks, even when code passes human audits.
Strengthening Defensive AI and Ecosystem Safeguards
Beyond the public EVMbench, OpenAI is expanding the private beta of Aardvark, an autonomous ‘white-hat’ security research agent. Aardvark operates in a continuous loop, monitoring code commits in real-time, hypothesizing attack vectors, verifying vulnerabilities in a secure sandbox, and even generating pull requests with proposed fixes without human intervention.
Coinciding with the release of EVMbench, OpenAI has also committed $10 million in API credits through its Cybersecurity Grant Program. This funding is specifically targeted at open-source projects and critical infrastructure to support defensive research in the blockchain space. Furthermore, OpenAI has open-sourced the EVMbench dataset and evaluation harness, encouraging researchers and developers to contribute new vulnerability patterns and refinement metrics, fostering a collaborative approach to cybersecurity.
The launch of EVMbench marks a pivotal moment in blockchain security, establishing a standardized framework to measure AI’s evolving capabilities in a domain where the stakes are continuously rising. The observed disparity between AI’s offensive and defensive prowess underscores the urgent need for continued investment in adversarial training environments to ensure that defensive AI technologies can evolve faster than their malicious counterparts, safeguarding the future of decentralized finance.