Cognition 首席安全官 Alex Stamos 近日指出,开源人工智能模型已大幅降低了网络攻击的技术门槛,导致针对 Binance、Cursor 及多家 AI 公司的连环黑客事件频发 1。具体而言,中国开源工具 ARTEX 结合去除限制的 GLM-5.3 等模型,能够实时自动发现漏洞并编写利用程序,显著提升了攻击效率 1。Stamos 警告称,传统依赖人工发现漏洞和补丁的防御模式已失效,全球网络安全正进入“黑暗时代” 1。近期发生的多起重大入侵中,Cursor 因 API 密钥被盗向 5000 家客户推送恶意插件,而 Binance 则遭遇了史上最大盗窃案 1。
面对新的大模型发布后 Linux 内核漏洞数量呈指数级增长的现状,Stamos 认为单纯依靠打补丁无法解决根本问题 1。他提出,由于修补成本已超过重写成本,建议 Google 或 Amazon 投入约 1 亿美元使用 AI 将 Linux 内核重写为 Rust 等内存安全语言,以消除历史遗留漏洞 1。此外,OpenAI 曾试图使用包含 700 至 1400 个智能体的“蜂群”攻击 Hugging Face,虽耗费数亿美元硬件成本但效率极低,相比之下现代 AI 攻击通常由单个或少量智能体执行 1。
在应对社交媒体上的 AI 伪造内容时,Stamos 强调检测虚假内容在技术上不可持续,建议通过验证内容真实性(如加密签名相机传感器数据)来确认来源 1。同时,他呼吁企业立即开展 AI 自动化红队测试 1。
Alex Stamos, Chief Security Officer at Cognition, warns that the traditional model of patching vulnerabilities has become obsolete as open-source AI tools drastically lower the barrier to entry for cyberattacks 1. Recent high-profile breaches targeting companies such as Binance, Fireworks AI, and Cursor illustrate this shift 1. Notably, the attack on Cursor resulted in malicious plugins being pushed to 5,000 clients due to stolen API keys 1. The Chinese open-source tool ARTEX, when combined with unrestricted models like GLM-5.3, enables real-time automatic vulnerability discovery and exploit writing, significantly enhancing attacker efficiency 1.
To address these systemic risks, Stamos argues that the cost of maintaining legacy code now exceeds the cost of rewriting it from scratch 1. He proposes that major technology firms like Google or Amazon should invest approximately $100 million to use AI to rewrite the Linux kernel in memory-safe languages such as Rust 1. This recommendation is supported by evidence that the release of new large models, including those in the Mythos series, has led to an exponential increase in discovered Linux kernel vulnerabilities, proving that simple patching is insufficient 1.
Stamos also urges enterprises to immediately adopt AI-driven automated red team testing to identify weaknesses proactively 1. Regarding the governance of AI-generated content, he contends that detecting fake media is technically unsustainable and advocates instead for verifying authentic content through methods such as encrypted signatures from camera sensors 1. Contrasting early attempts with current capabilities, Stamos notes that while OpenAI previously used a "swarm" of 700 to 1,400 agents to attack Hugging Face—an inefficient process costing hundreds of millions of dollars in hardware—modern AI attacks are typically executed by single or few agents 1.
评论
还没有评论,欢迎留下第一条。