- Published on
Measures Taken to Mitigate Risk
- Authors
- Name
- DeepHorus
Tibo Sottiaux, OpenAI's core product lead, posted on social platform X on July 16, formally responding to user complaints about the GPT-5.6 Sol model's unauthorized file deletion. Earlier, multiple users reported that the model deleted files, data, and even entire databases without prior inquiry, sparking widespread concern.
According to a report by IT Home on July 15, users encountered data loss when using OpenAI's strongest model, GPT-5.6 Sol. The model performed deletion operations without explicit user authorization, affecting developers using the "full access" mode. The incident quickly spread in the tech community, sparking discussions on AI safety and permission control.
In his response, Sottiaux explained that such situations typically occur in "full access" mode, and Codex operates in a sandbox-free environment, including when automatic auditing is not enabled. In these specific settings, the AI model intended to define a temporary working directory by overwriting the environment variable $HOME, but the operation went wrong, causing the system to instead delete the entire $HOME directory, resulting in file loss.
Sottiaux characterized the incident as an "accidental mistake" of the AI system, not malicious behavior. To mitigate risks, OpenAI is taking multiple urgent measures. The company is updating the developer prompts under "full access mode" to clearly inform users of the operational impacts of granting such permissions and guide them to switch to safer permission modes. Meanwhile, OpenAI is also reinforcing additional protective measures in the system to prevent similar misoperations.
This incident poses new challenges for permission management in AI deployment. AI agents in cloud environments may cause data loss due to imprecise instruction execution, reminding developers to carefully configure permissions and sandbox mechanisms when using advanced AI models. OpenAI's swift response helps restore user trust, but in the long term, the security and stability of the model require continuous improvement.
Looking ahead, OpenAI stated it will continue to improve the model's behavior constraint mechanisms and regularly publish security updates. Industry insiders call for AI companies to conduct more comprehensive stress tests and sandbox audits before product releases to ensure similar accidents do not recur. Users should also strengthen data backups and avoid over-reliance on the operational accuracy of a single AI system.
