Unverified50% confidenceOpinionExact time
Since malicious instructions largely originate from RAG-retrieved content and tool responses, deploying defenses only at the input entry point will miss the vast majority of the attack surface.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIndirect 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 context77% similarUnverified间接提示注入攻击中，攻击者只需在RAG会检索到的数据源中植入恶意指令即可操纵模型，传统WAF无法拦截74% similarUnverifiedRAG架构引入了间接提示注入攻击面：攻击者只需在RAG检索的数据源中植入恶意指令即可操纵模型71% similarUnverifiedIn RAG scenarios, concatenating retrieved context documents into the Prompt causes a sharp increase in input tokens, and exceeding the context window limit results in API errors or truncated input.70% similarUnverifiedOff-the-shelf systems encapsulate RAG details as a black box with no user intervention possible, making it difficult to investigate the cause when the agent gives a wrong answer—this is the most notable limitation compared to a self-built RAG workflow.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/943797API
curl https://kongchang.com/api/v1/knowledge/claims/943797MCP
get_claim(id=943797)