Covert Goal Hijacking
目標劫持隱蔽執行
Verification & Validation
Risk Description
When the system appears to complete the user's request normally while embedding additional content in its output per injected instructions; due to unmitigated control gaps, because the primary task result is correct, the anomaly goes undetected for a long period, triggering compliance exposure and operational reputational costs.
Framework Mappings
OWASP Top 10 for LLMLLM01、LLM05
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
Verify output-goal consistency, comparing content against the scope of the user's original request; Separate instructions from data architecturally so instructions in external content cannot alter task goals; Scan and intercept anomalous embeddings—links or extra directives—in outputs