📄 中文摘要
AI 执行风险是指系统在执行已批准的操作时，所依据的上下文信息可能已经发生变化，从而导致执行结果不再有效。在许多 AI 和机器学习系统中，决策是在上游进行的，而执行则是在后续阶段进行的。这种推理与执行之间的差距是导致问题的根源。在实际的软件工程中，AI 执行风险表现为多种形式，例如代理跳过步骤却仍然报告成功、工作流在过时数据上运行、系统在错误的时间执行正确的操作等。这些问题表明，AI 治理在执行阶段存在明显的缺陷，需要引起重视。
什么是 AI 执行风险？为什么 AI 治理在执行边界上失败
📄 English Summary
What Is AI Execution Risk? Why AI Governance Fails at the Execution Boundary
AI execution risk refers to the situation where a system performs an action that was previously approved, but the context has changed by the time it is executed, rendering the action invalid. In many AI and machine learning systems, decisions are made upstream and executed later. This gap between reasoning and execution is where failures occur. In real-world software engineering, AI execution risk manifests in various ways, such as agents skipping steps yet reporting success, workflows running on outdated data, and systems executing correct actions at the wrong time. These issues highlight significant flaws in AI governance at the execution boundary that need to be addressed.