Can AI Secure the Blockchain? Measuring Frontier Models on the EVMbench Framework скачать в хорошем качестве

Can AI Secure the Blockchain? Measuring Frontier Models on the EVMbench Framework

Can AI agents protect billions in on-chain assets? We analyze OpenAI’s EVMbench, a revolutionary benchmark evaluating how frontier models detect, patch, and exploit smart contract vulnerabilities. The Deep Dive EVMbench introduces a comprehensive evaluation suite comprising 120 high-severity vulnerabilities sourced from 40 real-world repositories. The methodology centers on three distinct operational modes: Detect (auditing via recall), Patch (modifying code while maintaining functionality), and Exploit (interacting with a live, local Ethereum instance via an RPC endpoint). To ensure environmental isolation and realism, the researchers implemented a Rust-based re-execution framework that replays agent transactions against a sandboxed Anvil blockchain. The results demonstrate a substantial leap in capability; models like GPT-5.3-Codex and Claude Opus 4.6 are now capable of executing end-to-end exploits against live blockchain instances. However, the study identifies a fundamental bottleneck: vulnerability discovery. When agents are provided with "hints" regarding the specific code mechanisms to examine, their success rates in patching and exploitation jump dramatically, suggesting that search and discovery—not technical repair—is the primary limitation for current frontier models. As smart contracts continue to manage massive financial resources, these findings offer a significant perspective on both the defensive potential of AI auditing and the emerging risks of automated exploitation. Academic Integrity Section This episode is a summary produced for educational and informational purposes by SciPulse. While we strive for technical precision, viewers should consult the original peer-reviewed paper by Wang et al. (OpenAI, Paradigm, OtterSec) for full data sets, oracle solutions, and methodology specifications. • Original Paper: [https://cdn.openai.com/evmbench/evmbe...] • Official Code & Data: [https://github.com/openai/frontier-evals] #SciPulse #AIResearch #SmartContractSecurity #OpenAI #BlockchainTech #Cybersecurity #EVMbench #MachineLearning #CryptoSecurity #AIModelEvaluation #STEM #EthSecurity #Ethereum #VulnerabilityResearch #FormalVerification