Unverified50% confidenceFactExact time
提示词注入的核心原理在于LLM无法从根本上区分系统指令和用户输入，因为Transformer架构的注意力机制对所有输入token一视同仁，系统提示词与用户输入的token在数学上处于同一向量空间
1
Sources
50%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
VerifiedTransformer-based large language models use attention mechanisms to uniformly encode system prompts, user inputs, and externally retrieved content as homogenized token sequences, making it impossible for the model to distinguish trusted instructions from untrusted data at the semantic level.78% similarUnverifiedThe root cause of prompt injection lies in the fact that LLM architecture encodes system prompts, user inputs, and externally retrieved content uniformly into the same continuous token sequence, making it impossible for the model to distinguish trusted instructions from untrusted data at the semantic level.77% similarUnverified提示注入攻击的根源在于LLM在推理时无法从语义层面可靠区分「系统指令」与「用户数据」，两者均以自然语言形式存在于同一上下文窗口中75% similarVerifiedThe memory of large language models fundamentally relies on the attention mechanism of the Transformer architecture, and its effective information capacity is limited by the number of tokens in the context window.74% similarUnverified自注意力本身对顺序不敏感，打乱词语顺序注意力权重计算结果相同，因此需要引入位置编码；原始Transformer论文使用正余弦函数编码71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/741045API
curl https://kongchang.com/api/v1/knowledge/claims/741045MCP
get_claim(id=741045)