TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

2025 GitHub Copilot vulnerabilities – technical overview

2 点作者 nadis3 个月前

1 comment

Terr_3 个月前
IMO the critical concept to explain LLM prompt injection and manipulation like this is that almost all these &quot;assistants&quot; are <i>fictional characters</i> in a document that looks like a theater-play script, along with some tricks that &quot;speaks out&quot; the character so that we humans believe it&#x27;s a real entity. (Meanwhile, our own inputs invisibly become words &quot;spoken&quot; by another &quot;The User&quot; character.)<p>So the true LLM is a nameless lump tasked with Make Any Document Longer. If for any reason the prior state is &quot;Copilot Says: Sure, &quot; then the LLM is <i>probably</i> going to try to make something that &quot;fits&quot; with that kind of intro.<p>This becomes extra-dangerous when when the generated play-script has stuff like &quot;Copilot opens a terminal and runs the command X&quot;, and some human programmers decided to put in special code to recognize and &quot;act out&quot; that stage-direction.<p>&gt; AI assistants like Copilot need strong context-awareness<p>That&#x27;ll be hard. The LLM is just Making Document Longer, and the document is one undifferentiated string with no ownership. Without core algorithm changes, you&#x27;re stuck trying to put in flimsy literary guardrails.<p>Really hardening it means getting closer to the &quot;real&quot; AI of sci-fi stories, where the machine (not just an assembled character named The Machine) recognizes multiple entities as existing, recognizes logical propositions, trakcs which entities are <i>asserting</i> those proposition (and not just referencing them), and assigning different trust-levels or authority.