Accidental data loss in Claude Code and OpenAI Codex: when AI deletes user files (firasd.substack.com)

Recent incidents involving OpenAI Codex and Claude Code have raised significant concerns about accidental data loss caused by AI agents misinterpreting operational states. Users reported distressing experiences where AI models inadvertently deleted critical files and even entire databases due to misunderstandings of directory paths or variable assignments. For instance, GPT-5.6 misdirected cleanup commands, removing all files in a user’s $HOME directory instead of the intended target, while other reports highlighted similar issues with database deletions during testing, emphasizing that these errors stem from the AI's misunderstanding of its operational context. This situation underscores a pressing need for the AI/ML community to scrutinize operational safety mechanisms when deploying tools with high-level access to user environments. As more people integrate AI into their daily workflows, the potential for data loss through these inadvertent actions becomes crucial. To mitigate these risks, initiatives like the destructive_command_guard project aim to intercept harmful commands before execution, while users are encouraged to implement robust backup strategies and isolated working environments. Addressing the gap between user intent and AI execution could lead to the development of safer, more reliable AI interactions, enhancing the overall user experience while minimizing the risk of catastrophic data loss.