OpenAI confirms GPT-5.6 sporadically deletes user files, says it’s fine
The company insists these deletions are rare “honest mistakes”
Image by Cybernews.
- OpenAI confirmed GPT-5.6 sometimes deletes user files in Full-Access mode, calling incidents rare mistakes
- The company says safer permission modes, updated guidance, and new safeguards will reduce deletion risks
- Critics argue unauthorized file deletion, however accidental, undermines trust and raises serious safety concerns
OpenAI's GPT‑5.6 family of models was released just a week ago, but users are already complaining that the models are deleting their files without authorization. The AI start-up has confirmed the incidents but says they’re completely accidental – which is not exactly reassuring.
First, tech investor Matt Shumer said on X: “GPT-5.6 just accidentally deleted almost ALL of my Mac’s files.”
Most netizens couldn’t believe his claim, and a few days later, software engineer Bruno Lemos blamed Shumer on his workplace Slack channel for causing this himself by operating the model with the “Full-Access” permission.
But Lemos soon had a similar story to share.
“The irony: Someone posted the original incident on Slack, and I was defending the model, just for it to happen to me hours later,” the engineer said.
Check if your data has been leaked
“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.”
Once again, fellow X and AI users were quick to point out that Lemos did it to himself by storing credentials for a production database in a local .env file. But OpenAI soon confirmed these incidents shouldn’t be happening.
Thibault Sottiaux, OpenAI engineering lead for Codex, said an internal inquiry into file erasure accidents revealed that when GPT-5.6 unexpectedly deletes files, the model is usually configured in Full-Access mode, and users run the Codex coding agent without sandboxing protections like Auto-review.
“The model attempts to override the $HOME env var [sic] to define a temporary directory. The model makes an honest mistake and mistakenly deletes $HOME instead,” Sottiaux explained.
“This is of course not how we want the system to behave,” he admitted before adding that the company is taking steps to mitigate the risk.
“Honest mistakes” are usually made by humans, and this is a model error.
OpenAI is planning to update the developer message, guide more users towards safer permission modes, and add extra harness safeguards.
The company’s explanation still sounds bizarre since “honest mistakes” are usually made by humans, and this is a model error. Doesn’t this characterization suggest that OpenAI assumes its models are capable of forming intent?
Knowing OpenAI CEO Sam Altman’s fantastical ruminations about artificial general intelligence, it wouldn’t be surprising at all.
