S0126RS3-T10-S0126-Z · Full risk code

Retrieval Injection Hijacking Output

檢索結果注入劫持輸出

Verification & Validation
Risk Description

When an attacker places a document containing hidden instructions in a public location the organization automatically indexes; due to unmitigated control gaps, when users query the relevant topic, the system retrieves the document and alters its response direction according to the embedded instructions, entirely without the user's awareness, triggering compliance exposure and operational reputational costs.

Framework Mappings

OWASP Top 10 for LLMLLM01、LLM08
MITRE ATLASAML.T0051
NIST AI 600-1Information Security
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 2

Risk Treatment & Implementation Guidance

Filter retrieval results, stripping instruction-like content from documents before generation; Control auto-indexing scope so documents in public locations are verified before entering the knowledge base; Periodically verify knowledge-base integrity, detecting documents with hidden instructions