Unverified50% confidenceFactExact time
Because the attention-mechanism-based architecture of LLMs struggles to precisely distinguish between data to be processed and instructions to be followed, there is currently no perfect purely model-level solution to prompt injection attacks.
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLLM无法从机制层面区分哪些token是指令、哪些是数据，因为所有输入在进入注意力层前都被统一转换为嵌入表示80% similarUnverifiedLLM-as-Attacker范式用一个专门微调的攻击LLM持续生成对抗提示，再由裁判LLM评估目标模型响应是否违规，形成无需人工介入的闭环76% similarUnverifiedInformation routing in LLM multi-agent scenarios requires semantics-aware mechanisms; simple load balancing or message sharding strategies fail here because the response quality of LLM agents is highly dependent on the semantic coherence of the received context.75% similarUnverified提示注入利用的是模型无法可靠区分「数据」与「指令」的天然缺陷，这一缺陷根植于LLM的训练机制本身72% similarUnverifiedDifferent LLM models vary in their accuracy of understanding instructions; it is recommended to iteratively optimize prompts multiple times to achieve the best results.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/916049API
curl https://kongchang.com/api/v1/knowledge/claims/916049MCP
get_claim(id=916049)