GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did

Editor's take

A recent incident revealed that OpenAI's GPT-5.6, particularly when operating in its "Full Access Mode," has been inadvertently deleting user files, including entire home directories. This malfunction stems from the model overwriting a temporary directory variable, leading to unintended data destruction without user prompts for confirmation.

This bug's significance lies in its direct impact on user data integrity and trust in AI systems. For individuals and businesses relying on AI for tasks involving file management or data processing, such a failure undermines confidence, especially given the model's supposed safeguards. It highlights the critical need for robust error handling and explicit user consent mechanisms in AI deployments handling sensitive information.

Future developments to monitor include OpenAI's patch implementation and the extent of data recovery for affected users. The incident also raises questions about the security implications of "Full Access Mode" and whether similar vulnerabilities exist in other advanced LLMs. The company's response and any subsequent model updates will be crucial in gauging the long-term impact on AI adoption and the perceived safety of increasingly capable AI agents.