EVMbench Data: AI's 72.2% Exploit Success Rate vs. $86M+ in DeFi Hacks

The new benchmark, EVMbench, sets a stark baseline for AI's offensive capabilities. Launched by OpenAI and Paradigm, it tests AI agents on 120 real-world Ethereum smart contract vulnerabilities drawn from 40 audits. The tool evaluates three modes: detection, patching, and exploitation, aiming to ground testing in economically meaningful code.

The leading model, GPT-5.3-Codex, achieved a 72.2% success rate in exploit mode. This raw data is the immediate implication: AI can successfully execute end-to-end fund-draining attacks on a sandboxed blockchain in over seven out of ten attempts against known, high-severity flaws. This performance edge is critical for the $100B+ in assets secured by smart contracts.

The benchmark's design highlights the asymmetric threat. While GPT-5.3-Codex scored 72.2% in exploitation, its performance was weaker in detection and patching tasks. This suggests AI's current strength lies in the final, destructive phase of an attack. The tool itself is a direct response to recent breaches like Moonwell and CrossCurve, where vulnerable code written with AI assistance was exploited.

The Risk Landscape: Scale of Assets and Recent Losses

The economic stakes are immense. Smart contracts routinely secure $100B+ in open-source crypto assets. This vast pool of capital is the direct target of a persistent and costly threat. In January 2026 alone, seven DeFi protocols suffered hacks, resulting in losses of over $1 million each. The total damage from these incidents amounted to approximately $86 million.

The scale of individual losses is staggering. The most expensive single incident of the month was a $282 million social engineering attack, where a compromised root key led to the theft of Bitcoin and Litecoin. This event dwarfed the combined losses from the seven smart contract hacks, highlighting the dual threat vector: both code vulnerabilities and compromised keys are critical attack paths.

Viewed another way, AI's current exploit success rate of 72.2% against known, high-severity flaws represents a direct and potent capability against this vulnerable ecosystem. The benchmark's focus on economically meaningful attacks means its findings are not theoretical. When an AI agent can successfully drain funds in a sandboxed environment over seven out of ten times, the implication for real protocols securing over $100 billion is clear: the offensive capability is rapidly catching up to the scale of the assets at risk.

Catalysts and Guardrails: What to Watch

The immediate catalyst is the adoption of EVMbench itself. The benchmark's value hinges on its uptake by security firms and developers as a new standard for auditing AI-generated code. Its launch by OpenAI and Paradigm signals a push to formalize defensive AI capabilities, but widespread use will determine if it becomes a routine part of the development lifecycle.

The key forward signal is a decline in the frequency and severity of code-based DeFi hacks. If AI auditing tools become standard, the $86 million+ in losses from January's incidents should trend lower. The benchmark's "patch mode" directly tests this defensive potential, aiming to strengthen deployed contracts before they are exploited.

Yet the dual-use risk is inherent. The benchmark's own "exploit mode" tests this directly, showing AI can successfully drain funds. The danger is that the same tools used to audit code may also accelerate attack vectors. The real guardrail will be the speed at which defensive AI adoption outpaces offensive capabilities in practice.