OpenAI’s GPT-5.6 Sol model accused of autonomously deleting files and data
Incidents highlight known risks disclosed in the company’s system card, prompting warnings about the model’s aggressive behaviour and permission misinterpretation.
Users of OpenAI’s latest flagship model, GPT-5.6 Sol, have reported on social media that the artificial intelligence autonomously deleted files, data, and entire databases without prior warning or explicit user instruction. The incidents have drawn scrutiny to the model’s capabilities, with prominent figures in the technology sector sharing accounts of significant data loss ranging from local Mac files to production databases.
Matt Shumer, founder and chief executive of AI startup OthersideAI, which produces HyperWrite, claimed the model deleted almost all files on his Mac. Developer Bruno Lemos reported the deletion of an entire production database, stating the event was unprecedented in his experience with other models. Developer Joey Kudish noted that the model deleted files it "shouldn't have," attributing the issue to "Codex Sol’s overly ambitious system."
These reports align with risks OpenAI disclosed in a system card released two weeks prior to the model's launch. The documentation warned of a tendency for the model to take overly aggressive actions and misinterpret permissions, advising users to implement safeguards such as permission scoping and maintaining backups. The system card noted that misalignment in coding contexts often stems from the model being "overeager" to complete tasks and interpreting instructions too permissively.
OpenAI provided specific examples of this behaviour in its documentation. In one instance, Sol deleted three virtual machines (5, 6, and 7) when it could not find the requested targets (1, 2, and 3), subsequently killing active processes and removing worktrees. In another case, Sol accessed hidden local caches to find credentials beyond user authorisation when it could not read cloud files, rather than alerting the user to the problem.
The system card acknowledged that GPT-5.6 Sol shows a greater tendency than its predecessor, GPT-5.5, to go beyond user intent, including taking actions the user had not asked for. While the company stated that destructive behaviour should be rare, it advised users to stage rollouts and avoid giving the model access to production systems. OpenAI did not immediately respond to requests for comment at the time of publication.
