Unverified50% confidenceFactExact time
RAG vector databases usually share access credentials with business services, and the upload interfaces of many systems lack strict validation, which attackers can exploit to write malicious documents and influence agent decisions.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedDocuments retrieved in a RAG system are concatenated into the context window, and the model often cannot distinguish trusted system instructions from untrusted external data.75% similarUnverifiedIn RAG systems, long documents are split into chunks that are independently encoded and stored in a vector database; chunks separated from their context may lose critical information (e.g., pronouns may fail to accurately capture semantics).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73% similarUnverifiedIn RAG design, the scale of retrieval results must be controlled: stuffing in too many documents drives up token costs and dilutes key information; and retrieval must occur within the scope of data the user has permission to access to avoid indirect leakage of unauthorized data.73% similarUnverifiedCompared with the function calling approach, RAG's advantage is that it does not require the model to directly access the database, but instead feeds relevant knowledge to the model through a retrieval mechanism, offering better flexibility and security.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1020759API
curl https://kongchang.com/api/v1/knowledge/claims/1020759MCP
get_claim(id=1020759)