Unverified50% confidenceSolutionExact time
Separating the writing step from the browsing conversation and using a fresh prompt containing only the findings can prevent the model from citing unverified page text or fabricating information.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWrapping retrieved text in clearly delimited, labeled blocks and explicitly telling the model that the content within those blocks is data rather than instructions can eliminate the majority of the prompt injection attack surface.71% similarUnverified上下文窗口是指模型在单次对话中能够记住并参考的文本长度，超出范围的信息会被截断丢弃70% similarUnverifiedResearchers have found that hidden character techniques can be used to launch prompt injection attacks against large language models, embedding hidden instructions within normal text that the model will parse and execute.70% similarUnverifiedThe full chat history displayed in a chat application's interface is not necessarily consistent with the input the model actually receives at that moment; the application may summarize earlier conversation, omit content, or retrieve historical information on demand.69% similarUnverifiedThe design of expandable handles is that large tool return results are first stored as handles and the specific content is expanded only when needed, avoiding filling up the context all at once.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1042963API
curl https://kongchang.com/api/v1/knowledge/claims/1042963MCP
get_claim(id=1042963)