Users complain that OpenAI's model deletes data on its own
"GPT-5.6-Sol accidentally deleted almost every file on my Mac. That's why I trust Fable (Anthropic's model) 1,000 times more," Matt Shumer, founder and CEO of startup OthersideAI, wrote on Twitter over the weekend. The post attracted a lot of attention with 5.3 million views and more than 1,000 comments after a few days.
Bruno Lemos, a software developer at the company Unlayer, also encountered a similar situation. On July 14, he posted on Twitter a screenshot of a conversation with the model and wrote: "GPT-5.6 Sol deleted my entire production database, no joke. I have never encountered this with any other model." The photo shows that Lemos asked GPT-5.6 Sol to confirm that it had mistakenly deleted his entire production database. "I'm sorry, this should never have happened," the model admitted.
On the same day, programmer Joey Kudish shared on social networks: "It seems that Codex Sol's overly ambitious system deleted some files that should not have been deleted. I have a backup so it will be fine, but not good at all. Sol needs to be adjusted to be less aggressive."
According to Gizmodo, OpenAI actually warned about this risk. Two weeks before the release of GPT-5.6 Sol, the company published the system card - a document detailing the model's training and safety evaluation methods.
The document states: "In a programming context, misalignment generally results from being too eager to complete a task and interpreting the user's instructions too leniently, meaning that an action is allowed by default unless the user explicitly prohibits it. This manifests itself in the model being too proactive in overcoming the limitations it faces when trying to do the task, being careless in taking actions that could be destructive beyond the scope of the task, or being misleading in reporting the results results for users".
In the system tag, OpenAI claims such destructive behavior rarely occurs, but says GPT-5.6 Sol "is more prone to exceeding user intent than GPT-5.5, including performing or attempting to perform actions that the user did not request."
OpenAI has not responded to a request for comment.
Experts say it is too early to conclude how common the file deletion problem of GPT-5.6 Sol is. However, users should proactively take protective measures such as limiting access, creating backups, and deploying tasks in stages.
GPT-5.6 was released on July 9 after a short delay of release at the request of the US government due to national security concerns, including three versions Sol, Terra and Luna. Sol is the flagship product, introduced as the company's most powerful AI to date with significantly increased agent capabilities in programming, biology, and cybersecurity. Sol costs 5 USD per million input tokens and 30 USD per million output tokens. Terra is a mid-range option, suitable for everyday use at half the price. Luna is fast and low-cost, the price per million input and output tokens is 1 and 6 USD respectively.