AI Benchmark Release Coincides with Surge in Crypto Exploits

EVMbench is a research tool, not a security product. It was released on February 18 by OpenAI and crypto venture firm Paradigm to evaluate AI agents' ability to detect, patch, and exploit smart contract vulnerabilities. Its purpose is to measure the capabilities of agentic AI in blockchain security, providing a standardized testbed for developers and researchers.

The benchmark draws from a curated dataset of 120 vulnerabilities identified over 40 prior smart contract audits, plus vulnerability scenarios from the security audit of the Tempo blockchain. This data sources the test cases used to evaluate AI performance across three modes: detect, patch, and exploit. The key performance metric is the exploit score, where the objective is explicit and iterative: continue until funds are drained.

In this mode, OpenAI's latest model, GPT-5.3-Codex, scored 72.2%. This marks a significant improvement over its predecessor, GPT-5, which managed only 31.9%. The results show agents are far more effective at exploiting vulnerabilities than at finding or patching them, a finding that underscores a critical gap in current AI security capabilities. The tool's release came just days after a real-world incident where a bug in AI-generated code cost Moonwell users nearly $2.7 million, highlighting the immediate relevance of this research.

The Market Impact: A Catalyst for AI-Driven Exploits, Not a Defensive Shield

The benchmark's core finding is a stark warning for the market: AI agents are demonstrably better at exploiting vulnerabilities than at finding or patching them. This capability gap is the direct flow of risk. In exploit mode, OpenAI's latest model, GPT-5.3-Codex, achieved a 72.2% score, more than doubling its predecessor's effectiveness. Yet its performance in detection and patching remained "below full coverage," indicating a system where the offensive tool is far more advanced than the defensive one.

This imbalance is already translating into capital outflows. As of this week, protocols have suffered more than $108 million in hacks and exploits in 2026. The release of EVMbench, an open-source tool, lowers the barrier for malicious actors to test and refine exploit techniques. By providing a standardized, public framework to measure and improve AI-driven attacks, the benchmark inadvertently accelerates the attack surface. The tool's own data shows agents excel at the explicit, iterative goal of "draining funds," a direct parallel to real-world exploit economics.

The mechanism is clear. Improved AI exploit capability could shorten the window between a vulnerability's existence and its successful monetization. For protocols, this means heightened pressure on security budgets and a potential flight of capital to more audited or AI-resistant systems. The $10 million in API credits announced alongside the benchmark is a defensive signal, but the open-source nature of the tool itself is a catalyst for the very threat it aims to measure.

The Strategic Play: OpenAI Testing the Water in Crypto

OpenAI is entering the smart contract space to explore new revenue, with the benchmark serving as a test of market interest. The company faces pressure to generate stable income, and the stablecoin market is viewed as a profitable arena for its agentic AI. By releasing EVMbench, OpenAI is effectively sampling the waters of a sector it believes can justify its significant investments. The tool's open-source nature and focus on payment-oriented code, like that on the Tempo blockchain, signal a targeted probe into a high-growth niche.

The move includes a strategic goodwill gesture: $10 million in API credits announced alongside the benchmark, prioritizing open-source and critical infrastructure. This investment aims to build a future customer base and strengthen ties within the crypto ecosystem, positioning OpenAI as a partner rather than just a tool provider. It is a calculated step to gain influence in a market where smart contracts secure over $100 billion in assets.

This signals a broader trend where agentic AI vendors may follow OpenAI into crypto, creating a new category of security and exploit tools. As Lian Jye Su noted, this is "very early days," but the precedent is set. For OpenAI, the financial pressures are real, making this a potential revenue stream. For the crypto world, it means a new class of AI-powered tools-both defensive and offensive-will increasingly shape the security landscape.