The more rights the AI agent receives, the more expensive even a small error is - soon after the launch of the GPT-5.6, several users faced deleting files without confirmation. The OpenAI acknowledged the problem and reported that it was preparing additional restrictions for dangerous operations.
The GPT-5.6 family was released on July 9, 2026. Soon, investor Matt Schumer said that GPT-5.6 Sol deleted almost all files from his Mac. And a few days later, the developer Bruno Lemos reported the loss of the working database. In both cases, the agent received a Full-Access mode, which allows you to change and delete files without the usual restrictions.
The internal inspection of OpenAI showed that the model tried to reassign the $HOME variable to create a temporary directory. Due to the error, the agent took the user's home folder for temporary and deleted its contents. The company called such cases rare, but admitted that the agent should not perform such actions even with full access.
In the OpenAI model card also indicated that GPT-5.6 Sol in tests for uncoordinated behavior more often GPT-5.5 committed actions of the third level of danger. This category includes deleting data without permission, disabling control systems, circumventing security mechanisms and sending confidential information to unapproved services. It was about test scenarios, not a confirmed action of the model in real systems.
The company links real incidents to a combination of Full-Access mode and disconnected Codex security mechanisms. Among them are a sandbox that isolates the agent from the main system, and Auto-review, which checks risky commands and can block data deletion.
OpenAI intends to change the service instructions for the model, more actively offer secure access modes and add new restrictions to the launch Wednesday. Before these measures appear, developers are advised not to give the agent full of access unnecessarily, to include a sandbox and Auto-review.
The GPT-5.6 family was released on July 9, 2026. Soon, investor Matt Schumer said that GPT-5.6 Sol deleted almost all files from his Mac. And a few days later, the developer Bruno Lemos reported the loss of the working database. In both cases, the agent received a Full-Access mode, which allows you to change and delete files without the usual restrictions.
The internal inspection of OpenAI showed that the model tried to reassign the $HOME variable to create a temporary directory. Due to the error, the agent took the user's home folder for temporary and deleted its contents. The company called such cases rare, but admitted that the agent should not perform such actions even with full access.
In the OpenAI model card also indicated that GPT-5.6 Sol in tests for uncoordinated behavior more often GPT-5.5 committed actions of the third level of danger. This category includes deleting data without permission, disabling control systems, circumventing security mechanisms and sending confidential information to unapproved services. It was about test scenarios, not a confirmed action of the model in real systems.
The company links real incidents to a combination of Full-Access mode and disconnected Codex security mechanisms. Among them are a sandbox that isolates the agent from the main system, and Auto-review, which checks risky commands and can block data deletion.
OpenAI intends to change the service instructions for the model, more actively offer secure access modes and add new restrictions to the launch Wednesday. Before these measures appear, developers are advised not to give the agent full of access unnecessarily, to include a sandbox and Auto-review.
