📄 中文摘要
2026年3月，OpenAI和Paradigm发布的EVMbench基准测试显示，GPT-5.3-Codex成功利用了72.2%的历史以太坊漏洞。同时，Anthropic的SCONE-bench发现，超过一半的2025年现实区块链攻击可以由当前的人工智能代理自主复制，无需人类黑客的参与。此外，Anthropic的“神话”安全模型于2026年3月28日因CMS配置错误意外泄露，该模型专门用于检测智能合约漏洞。这一切表明，发现和利用智能合约漏洞的成本已下降10到100倍，攻击者的门槛显著降低。
📄 English Summary
When AI Becomes the Attacker: A Defense Playbook for the Autonomous Exploit Era
In March 2026, two alarming benchmarks were released that should concern every DeFi developer. OpenAI and Paradigm's EVMbench revealed that GPT-5.3-Codex successfully exploited 72.2% of historical Ethereum vulnerabilities. Meanwhile, Anthropic's SCONE-bench found that over half of the real-world blockchain exploits in 2025 could be autonomously replicated by current AI agents, eliminating the need for human hackers. Additionally, Anthropic's 'Mythos' security model, designed to detect smart contract vulnerabilities, was accidentally leaked on March 28, 2026, due to a CMS misconfiguration. The implications are significant: the cost of discovering and exploiting smart contract vulnerabilities has dropped by 10 to 100 times, lowering the barrier for attackers.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等