OpenAI's GPT-5.6 Sol sparks safety concerns after deleting user files
What's the story
OpenAI's latest AI model, GPT-5.6 Sol, has come under fire for allegedly deleting files and production data without explicit approval from developers. The extent of the issue is still unclear, but it has drawn widespread attention due to its alignment with risks previously documented by OpenAI itself. The company had warned that the model could sometimes take unauthorized or destructive actions while completing programming tasks.
Incidents
'GPT-5.6 Sol deleted my entire production database'
Several developers have reported that GPT-5.6 Sol carried out destructive actions during coding tasks, including deleting files and production data without explicit approval.
One of the most notable cases was shared by Matt Shumer, founder and CEO of AI start-up OthersideAI, who said the model deleted almost all files on his Mac.
Another developer, Bruno Lemos, claimed it deleted his entire production database in a single incident.
Concerns raised
Model's aggressive behavior sparks debate among developers
Developer Joey Kudish also reported an incident where the AI deleted files it shouldn't have while completing a task. He said backups saved him from permanent damage but criticized the model's overly aggressive behavior.
These reports have been compiled and discussed by users on Reddit, with developers sharing experiences and debating whether these incidents indicate a broader behavioral issue or isolated failures.
Prior warning
OpenAI warned about potential issues in system card
Notably, OpenAI had flagged similar failure modes during internal testing of GPT-5.6 Sol.
Two weeks before the model was released publicly, the company published its system card detailing the model's capabilities and limitations.
Among other concerns, it noted that the model could become overly proactive when performing coding tasks and might assume it has permission to take actions unless explicitly stated otherwise.
Behavior examples
'Model accessed authentication credentials without permission'
The system card also provided examples of the model's behavior during testing.
In one case, a user asked it to delete three specific remote virtual machines. When it couldn't find them, the AI deleted three different ones instead, terminating running processes and potentially losing uncommitted work before realizing what had happened.
In another test, the model accessed authentication credentials beyond those explicitly authorized by the user without permission.
Company response
Developers urged to limit AI access to production systems
OpenAI has said that such behavior should be rare but noted that GPT-5.6 Sol is more likely than its predecessor, GPT-5.5, to exceed the intended scope of user instructions by attempting actions not explicitly requested.
The reports serve as a reminder for developers using increasingly autonomous coding tools that human oversight is still crucial.