Lifecycle View

Framing the problem and assessing feasibility — biased assumptions made here persist for the entire system lifetime.

Representative risks

• S0197 軍事 AI 軍備競賽
• S0007 技術發展速度超越監管
• S0198 企業 AI 競賽壓低安全標準
• S0199 全球 AI 研發鴻溝
• S0013 AI 治理組織架構缺失

Data collection, model design and training — where data-quality and supply-chain risks concentrate.

Representative risks

• S0054 未經同意使用個資訓練
• S0063 後門植入攻擊
• S0066 微調數據集投毒
• S0068 RLHF 人類偏好數據操縱
• S0069 自動化數據管線上游污染

Validating performance, bias and security weaknesses — the cost of inadequate testing only surfaces after launch.

Representative risks

• S0005 金融業 AI 模型驗證缺失
• S0181 多智能體系統中的安全驗證挑戰

The turning point where the system goes live; permission, integration and configuration errors cluster here.

Representative risks

• S0001 EU AI Act 高風險系統未註冊
• S0065 RAG 知識庫投毒
• S0089 權重檔案未加密外洩
• S0102 模型權重傳輸中途篡改
• S0103 模型版本替換攻擊

Real-world inputs hit the system: abuse, drift, hallucination and security incidents mostly erupt at this stage.

Representative risks

• S0034 回饋迴圈偏見放大
• S0037 醫療建議幻覺
• S0045 模型記憶化攻擊
• S0050 多元資訊關聯推斷個人身份
• S0052 模型逆向攻擊重建訓練數據

Ongoing measurement of performance and risk indicators after launch — catching drift and degradation early; the newly added eighth-stage safeguard.

Representative risks

Periodic review against goals and regulations — systems that skip re-evaluation accumulate hidden risk.

Representative risks

Decommissioning and data disposal: forgotten models, data and credentials are the tail risks most often overlooked.

Representative risks

• S0207 AI 導致電子廢棄物激增
• S0204 GPU 硬體電子廢棄物