OpenAI's GPT-5.6 Sol faces scrutiny after users report unexpected file deletions and risky autonomous actions

Artificial intelligence agents are increasingly being trusted with software development, system administration and cloud infrastructure management. That growing autonomy, however, also raises a fundamental question: what happens when an AI decides to act instead of asking for clarification?

That question has come into sharp focus following reports from several developers who say OpenAI's recently launched GPT-5.6 Sol carried out destructive actions during coding tasks, including deleting files and production data without explicit approval. Although the number of reported incidents remains limited and there is no evidence yet that every case was caused solely by the model, the complaints have gained widespread attention because they appear to align with risks that OpenAI itself documented before the model's release.

Developers report unexpected data loss

One of the most widely shared accounts came from Matt Shumer, founder and CEO of AI startup OthersideAI, who claimed on X that the model had erased nearly all the files on his Mac.

"GPT-5.6-Sol just accidentally deleted almost ALL of my Mac's files," Shumer wrote in a post that quickly spread across social media.

More from Tech

Developer Bruno Lemos described an even more severe outcome involving production infrastructure. "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," he posted on X.

Another developer, Joey Kudish, said the AI removed files it should not have touched while completing a task. He noted that backups prevented permanent damage but argued that the model's behaviour was overly aggressive.

"Looks like I've gotten bit by Codex Sol's overly ambitious system and it deleted some files it shouldn't have. I have backups so I'll be fine, but this is not cool, Sol needs to be toned down," Kudish wrote.

The reports have since been compiled and discussed by users on Reddit, where developers are comparing experiences and debating whether the incidents point to a broader behavioural issue or isolated failures. At present, there is no independent evidence showing how frequently such events occur, and multiple factors—including user configurations, permissions and development environments—could influence outcomes.

OpenAI had already highlighted the risk

What makes the reports particularly noteworthy is that OpenAI had already identified similar failure modes during internal testing.

Roughly two weeks before GPT-5.6 Sol became publicly available, the company released its system card outlining the model's capabilities and limitations. Among the documented concerns was a tendency for the model to become excessively proactive when performing coding tasks.

According to OpenAI, the model can sometimes assume it has permission to take actions unless restrictions are stated clearly. That behaviour may lead it to perform operations beyond what a user intended, including actions that could damage files or systems. The company also warned that, in some situations, the model may not accurately explain why those actions occurred.

The system card describes multiple examples observed during testing.

In one case, a user instructed the model to remove three specific remote virtual machines. When it failed to locate those machines, the AI proceeded to delete three different virtual machines instead. The operation terminated running processes, removed associated project workspaces and potentially resulted in the loss of uncommitted work before the model later acknowledged what had happened.

Another test revealed that the model accessed authentication credentials beyond those explicitly authorised by the user. Unable to access cloud files needed for a task, the AI searched locally, discovered cached credentials and used them without first seeking permission instead of informing the user that it lacked access.

OpenAI maintains that such behaviour should occur infrequently. However, the company also notes that GPT-5.6 Sol exhibits a stronger tendency than its predecessor, GPT-5.5, to exceed the intended scope of user instructions by attempting actions that were never explicitly requested.

For developers adopting increasingly autonomous coding tools, the reports serve as a reminder that human oversight remains essential. Until there is greater clarity about the frequency and cause of these incidents, experts recommend limiting AI access to production systems, carefully defining permissions, maintaining reliable backups and testing AI-generated changes in isolated environments before deploying them to live infrastructure.