Unverified50% confidenceCautionExact time
Any scenario in which instructions and untrusted data share the same context (such as prompt injection, RAG, or agents) carries the risk that the model may follow information it was told to ignore.
1
Sources
50%
Confidence
Long-term
Relevance
10/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified工具调用使攻击面从让模型说错话升级为让模型做错事，危害从信息误导上升为实质性的数据访问和系统篡改74% similarUnverifiedIn LLM agent systems involving retrieval-augmented generation (RAG), tool calling, and multi-step reasoning, a single minor prompt modification can cause a previously exposed risk path to disappear.73% similarUnverifiedAfter catching an exception, the error information should be returned to the large model verbatim rather than silently discarded, enabling the model to self-reflect, adjust its format, and retry, thereby improving agent robustness.71% similarVerifiedIndirect prompt injection conceals attack instructions within external documents, web pages, or database records retrieved by the model, and is activated when a RAG system introduces them into the context71% similarUnverified针对特定触发词的后门攻击可让模型在大多数情况下表现正常，仅在遇到特定输入时才输出预设错误信息，使常规评测基准难以捕捉71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/968651API
curl https://kongchang.com/api/v1/knowledge/claims/968651MCP
get_claim(id=968651)