精选理由
Ken Thompson的编译器安全论文，揭示了AI模型训练中可能存在的类似安全风险。
Ken Thompson在1984年发表的《Reflections on Trusting Trust》论文探讨了编译器安全漏洞。他描述了一种"有毒"编译器，能在编译自身时不留痕迹地植入后门。这种攻击方式与当前AI模型训练中的潜在安全风险高度相似，可能导致一个被污染的模型帮助训练下一代模型，同时清除所有污染痕迹。
原文 · Amjad Masad
Ken Thompson’s “Reflections on Trusting Trust” feels super relevant to AI. He bootstrapped a “poiso...
Ken Thompson’s “Reflections on Trusting Trust” feels super relevant to AI. He bootstrapped a “poisoned” compiler that left no traces of the poison in the source code because it compiled itself. You can imagine the same thing happening with models where one poisoned generation helps train the next while removing any traces of the poison. 💬 18 🔄 2 ❤️ 59 👀 4707 📊 22 ⚡