Unverified50% confidenceFactExact time
LLM本质上无法从架构层面区分被处理的数据和应当执行的指令，这类攻击极难从模型本身彻底根除
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified目前业界尚无完美的提示注入防御方案，指令与数据的根本性混淆是LLM架构层面的限制74% similarUnverifiedThe industry has not yet found a solution that fundamentally resolves prompt injection and jailbreak attacks, because LLMs are inherently unable to distinguish between instructions and data.73% similarUnverifiedThe stateless nature of LLMs (each conversation being independent) prevents them from maintaining design consistency across multiple interactions, which is a fundamental limitation at the technical architecture level.70% similarUnverifiedThe security risk of AI agent architecture lies in the fact that LLM decision-making is probabilistic rather than deterministic, lacks an independent verification layer, and the LLM acts as both policy maker and execution initiator, violating the principle of separation of duties.66% similarUnverifiedReAct的局限在于以模型本身为中心，缺乏对工具调用失败、模型幻觉、并发任务等工程级问题的系统性处理66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/535980API
curl https://kongchang.com/api/v1/knowledge/claims/535980MCP
get_claim(id=535980)