根据多条Mastodon帖子，数据投毒对大型语言模型（LLM）构成了重大且难以解决的安全风险。核心问题源于将LLM用作通用工具，超出了其最初为语言处理设计的目的。虽然投毒很容易执行，但认为完全防范LLM免受其害是不可能的，因此需要进行强大的威胁建模。 AI
影响 强调了LLM固有的安全漏洞，并着重指出需要先进的威胁建模作为缓解策略。
排序理由 多篇社交媒体帖子讨论了LLM因数据投毒而存在的固有安全风险，但并非来自主要来源或研究论文。
在 Mastodon — sigmoid.social 阅读 →
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
LLM Data poisoning, besides being extremely easy to do, is extremely difficult to fix. This because of the nature of LLs. Thry were built for language processing, but we use them for everything from coding to image recognition. #llm #ai m.youtube.com/watch?v=2uG8... bsky.app/prof…
LLMs aren‘t insecure because they have been badly trained, they are insecure because we are using them as a general purpose tool. Securing them is from data poisoning completely is impossible. Threat modeling is a necessary activity because of it. #llm #ai
LLM Data poisoning, besides being extremely easy to do, is extremely difficult to fix. This because of the nature of LLs. Thry were built for language processing, but we use them for everything from coding to image recognition. #llm #ai m.youtube.com/watch?v=2uG8... bsky.app/prof…