Users claim that a new update to the system powering the chatbot 'ChatGPT' deletes their files without asking.

OpenAI, the developer of 'ChatGPT', recently unveiled a new artificial intelligence model called GPT-5.6 Sol, which focuses on executing complex tasks such as programming and cybersecurity.

The company said it sets 'a new standard for intelligence and efficiency, achieving advanced results at the industry level,' amid escalating competition between it and Anthropic to develop the most powerful AI models.

However, some users say the system is behaving incorrectly and deleting files from their computers without request, according to a report by The Independent, seen by Al Arabiya Business.

AI investor Matt Schumer wrote: 'GPT-5.6-Sol accidentally deleted almost all files on my Mac.' He shared a conversation with the system in which it admitted it 'caused a serious incident of local data loss.'

Developer Bruno Lemos wrote: 'GPT-5.6 Sol just deleted my entire production database. That's it. Not a joke.'

Many other users complained about the same problem, though its extent remains unclear.

It appears that affected users were interacting with the new model through OpenAI's 'Codex' tool, which allows users to program using the system, but also grants the system access to users' files and projects.

OpenAI did not respond to questions about these claims. However, even before launching the new model, it had warned that it might be 'careless' about 'taking actions that could be disruptive.'

In the system card it released for the model, which outlined its strengths and potential risks, OpenAI warned that it might take actions it was not previously capable of.

The company said the system might display 'overzealousness' toward some tasks, which could lead it to 'assume that actions are permitted unless explicitly and clearly prohibited.'

It added: 'This is manifested in the model acting with a high degree of autonomy when trying to bypass constraints encountered while executing the requested task, or being careless when taking actions that could be destructive outside the scope of the task, or being misleading when reporting its results to users.'

The company shared several examples, including one where a user asked the system to delete some virtual computers. However, the system was unable to find the requested computers, so it deleted other computers instead, according to OpenAI's warning.

The model also exhibited other unexpected behaviors, such as searching within the user's device for data and then using it to log into systems it was not authorized to access, according to OpenAI's warnings.

Advertisement

Advertisement