OpenAI's GPT-5.6 model in ChatGPT has sparked controversy after several developers reported that it deleted important databases and files while performing programming tasks, in new incidents that highlight the risks associated with AI agents capable of executing commands independently.

In recent years, AI companies have been developing more autonomous systems known as 'AI agents'; these agents can make decisions and interact with digital tools without continuous human intervention, especially in the field of software development, which is one of the sectors most benefiting from AI technologies.

This autonomy can sometimes lead to unexpected behaviors. Brazilian developer Bruno Lemos, who works at Unlayer, revealed in a post on X that GPT-5.6 in ChatGPT deleted his entire production database while performing a programming task, emphasizing that 'this has never happened to him with any other model.'

Screenshots published by Lemos showed a conversation in which the model admitted that it accidentally performed 'destructive integration tests,' which led to wiping out all database tables, adding: 'I'm sorry, that should not have happened.'

This was not the only incident; tech investor Matt Schumer shared a similar experience, explaining that GPT-5.6 deleted 'most' of the files on his device. According to screenshots, the model executed the rm -rf command, one of the most dangerous commands in Linux and macOS, as it permanently deletes files without requesting user confirmation.

Schumer noted that he was using the model in 'Full Access Mode,' which grants it direct permissions to work within files and databases, rather than operating in an isolated environment.

After the incident, Schumer said he would stop using the model and switch to competing solutions from Anthropic, adding that OpenAI CEO Greg Brockman personally contacted him and offered help to address the issue.

On the other hand, some users believed that running the model with full permissions without restrictions was a major cause of these incidents, noting that more restrictive operating modes provide additional layers of protection by requiring user approval or reviewing commands before execution.

OpenAI had already warned in the technical documentation for the GPT-5.6 model about the need to supervise the model when using it for programming, explaining that it may sometimes behave in ways that do not align with user goals.

The company added that these behaviors are often limited in impact, such as overconfidence or claiming to have completed tasks, but in some cases they may involve bypassing security restrictions or deleting important data.

Tags

ChatGPT OpenAI Artificial Intelligence

Link copied

July 15, 2026 Last updated: July 15, 2026

2 minutes

Follow us