Unverified50% confidenceFactExact time
LLMs cannot reliably distinguish between 'system instructions' and 'user input,' which is the core logic of prompt injection attacks
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe core capability of LLMs is confined to the 'text input → text output' framework, making them unable to proactively read project files, execute system commands, or access external tools.68% similarUnverifiedSystem Prompt与用户输入本质上都是模型的输入token，模型并不存在独立的规则执行引擎来强制保证指令被遵守64% similarUnverifiedLLMs themselves have no memory capability whatsoever, and every response is as if the user is being encountered for the first time.64% similarUnverifiedCLAUDE.md is not a System Prompt — it is injected as a user message58% similarUnverifiedLLM 本身不具备直接操作文件系统、调用终端或感知代码库拓扑结构的能力，这些工程能力必须由 IDE 层封装实现57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/920812API
curl https://kongchang.com/api/v1/knowledge/claims/920812MCP
get_claim(id=920812)