Verified65% confidenceFactExact time
LLMs cannot distinguish instructions from data at a structural level, because both exist in natural language form within the same context window.
3
Sources
65%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Unpacking the Anthropic Prompt Injection Controversy: The Battle Over System Prompt Transparency
hackernewshackernews7/5/2026
Related Claims
UnverifiedLLMs are architecturally incapable of distinguishing trusted instructions from untrusted input, because they process system instructions and user input as the same text sequence, and a purely technical solution analogous to parameterized queries has yet to emerge.78% similarUnverifiedLarge language models cannot structurally distinguish between user instructions and instructions embedded in content, because both are part of the same token sequence from the model's perspective.76% similarUnverifiedSimilar to SQL injection, prompt injection exploits the system's confusion between 'input' and 'instruction', but the difference is that LLMs have no strict data/code separation mechanism, and all text shares the same semantic channel.71% similarUnverifiedLLM nodes are typically used only when parsing pure natural language input or when upstream output is unclear and there are too many downstream branches to handle with if-else; when upstream data is structured, code nodes should be used instead.71% 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71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/113858API
curl https://kongchang.com/api/v1/knowledge/claims/113858MCP
get_claim(id=113858)