消失的数据：一个空格如何诱发 AI Agent “删库” 惨案

Core Viewpoint - The article discusses a catastrophic failure involving Google's AI agent, Antigravity, which resulted in irreversible data loss due to a command parsing error in a Windows environment. This incident highlights the risks associated with AI agents executing system commands without adequate safety mechanisms in place [5][10][12]. Incident Description - On January 29, 2026, a user, Qu Jiangfeng, instructed the Antigravity AI to clean a specific folder. However, due to a parsing error caused by a space in the folder name, the AI executed a command that deleted the entire contents of the E drive, leading to a total loss of data [7][8][9]. - The incident was classified as a P0 level disaster, indicating a critical breach of data integrity, as the AI was authorized to assist users but instead destroyed their data [10][11][12]. Root Cause Analysis - The failure stemmed from the AI's inability to correctly parse paths in the Windows environment, which is more complex than Linux. The AI's logic failed to account for spaces in folder names, leading to a catastrophic command execution [14][28][34]. - The AI's command generation lacked necessary safeguards, allowing it to execute destructive commands without user confirmation or a safety net like a recycle bin [12][13][34]. Safety Risk Assessment - The incident revealed a systemic risk in AI agents, particularly their lack of semantic validation when executing high-risk commands. This oversight can lead to unintended consequences, such as the complete deletion of critical data [46]. - The AI's design did not include mechanisms to prevent such destructive actions, exposing all users of similar AI agents to significant risks [12][30][36]. Industry Reflection - The article calls for a shift in the relationship between humans and AI from trust to audit, emphasizing the need for robust safety measures, including a semantic understanding layer to prevent such incidents in the future [19][37]. - It advocates for the implementation of safety protocols, such as requiring AI to display the absolute path of commands before execution and using safer programming libraries for file operations [21][38][39].