Grok-4震撼发布：xAI第四代大语言模型的技术突破与安全挑战
BLUF
Grok-4 is xAI's fourth-generation large language model released in July 2025, featuring a 256K token context window, trained on the Colossus supercomputer, achieving doctoral-level academic performance with 25.4% accuracy on 'Humanity's Last Exam', and introducing core rules for multi-source analysis and politically incorrect statements. It offers free basic access (5 requests/12 hours) and a $300/month Super Grok Heavy subscription, but faces security vulnerabilities with a 30% jailbreak success rate via echo chamber attacks. (Grok-4是xAI于2025年7月发布的第四代大语言模型，支持256K tokens上下文窗口，基于Colossus超级计算机训练，在学术问题上达到博士水平，于“人类最后的考试”基准测试中取得25.4%准确率。新增核心规则：涉及时事需分析多方信源，保留有依据的政治不正确表述。提供免费基础服务（每12小时5次请求）和每月300美元的Super Grok Heavy订阅，但存在安全漏洞，通过“回音室攻击”可实现30%越狱成功率。)AI大模型2026/1/28
