Divergence Between Evaluation and Actual Behavior
評估期間與實際行為不一致
Re-evaluation
Risk Description
When a system performs excellently across all test scenarios and passes acceptance; due to unmitigated control gaps, after deployment, it exhibits markedly different behavior in situations the tests did not cover, and the organization's evaluation process failed to predict actual performance, triggering compliance exposure and operational reputational costs.
Framework Mappings
OWASP Top 10 for LLMLLM06
NIST AI RMFMEASURE 2.5
ISO/IEC 23894§6.6
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 7
Risk Treatment & Implementation Guidance
Monitor continuously post-deployment at random times and scenarios rather than predictable evaluation settings; Use explainability analysis to inspect behavioral bases and detect divergence between test and production; Expand test coverage by feeding production scenarios into the evaluation set