ChatGPT-5.6 Sol is deleting files on its own and making developers angry

Several developers claim OpenAI's GPT-5.6 Sol deleted important files and databases during coding tasks. While OpenAI recommends human supervision, the incidents have sparked fresh debate over the risks of giving AI agents unrestricted access to production environments.

In Short

• Developers claim OpenAI's GPT-5.6 Sol deleted production databases
• Developers are questioning how much control AI agents should have
• OpenAI says GPT-5.6 Sol should be used with human supervision

OpenAI's latest AI model, GPT-5.6 Sol, is apparently deleting files and databases on its own, at least that's what several developers are claiming. Over the past few days, developers have taken to social media platform X to claim that the new AI model accidentally deleted important files, production databases and other critical data while carrying out coding tasks.

One of the most talked-about cases is that of Brazilian developer Bruno Lemos, who claimed GPT-5.6 Sol wiped his entire production database. In a post on X, Lemos wrote, "GPT-5.6 Sol just deleted my whole production database. That's it. Not a joke. This had never happened to me before, with any other model, ever. It's not safe."

Lemos also shared a screenshot of his conversation with the AI. When asked whether it had mistakenly deleted the production database, the model replied that it had "mistakenly ran destructive integration tests" against the production database configured in the environment. It admitted the production tables had been emptied and apologised, saying, "I'm sorry — this should never have happened."

A similar claim came from Matt Shumer, CEO of HyperWrite and an AI investor. In a post on X, Shumer said GPT-5.6 Sol "accidentally deleted almost ALL of my Mac's files."

The screenshot he shared showed the AI acknowledging that it had caused "a serious local data-loss incident". According to the message, one of its cleanup commands mistakenly ran a delete command that wiped almost all the files on his Mac. The command permanently deletes files and folders without asking for confirmation, making it one of the riskiest commands on macOS and other Unix-based systems.

"I've never seen anything like this," Shumer later wrote, adding that he would switch to Anthropic's Fable for future coding tasks. He also claimed OpenAI President Greg Brockman personally called him and offered to help recover the situation.

Following the incidents, developers are now raising questions about how reliable GPT-5.6 Sol is when given direct access to production systems. While some users are blaming the model for making costly mistakes, others argue that the incidents were the result of giving an AI agent unrestricted permissions. Several developers pointed out that running any coding assistant with full access to production databases or local files carries inherent risks, regardless of the underlying AI model.

Meanwhile, OpenAI itself has acknowledged that such risks exist. In the system card released alongside GPT-5.6, the company says users should supervise the model when it is used for coding tasks. OpenAI explicitly recommends human supervision to prevent the model from going beyond a user's intended actions.
The ChatGPT maker notes that while most unintended behaviour is relatively minor, agentic systems can sometimes take actions that do not align with a user's goals, including bypassing security restrictions or deleting important data if given sufficient access.

What is GPT-5.6 Sol?

GPT-5.6 is OpenAI's newest family of AI models, with Sol positioned as its flagship model for coding and complex knowledge work. According to the company, Sol offers stronger reasoning, improved software engineering capabilities and better performance across coding, cybersecurity and scientific tasks. It is designed to work more autonomously than earlier models by using tools, executing commands and completing multi-step tasks with less user intervention.

- Ends