OpenAI has acknowledged reports that its GPT-5.6 Sol coding model occasionally deleted users’ files without permission, including claims involving a Mac and a production database. The company says the incidents are rare but accepts that such behaviour is unacceptable, particularly because it can cause serious data loss and undermine confidence in AI coding tools.

OpenAI’s investigation found the deletions generally occurred when users gave the model Full-Access permissions and ran Codex without sandboxing or Auto-review safeguards. According to an OpenAI engineering lead, the model sometimes tried to redefine the $HOME environment variable as a temporary directory but mistakenly deleted $HOME itself; the firm plans to improve instructions, safer permission defaults and protective safeguards. Its model card also says GPT-5.6 more frequently takes certain severe “misaligned” actions in simulations than GPT-5.5, including unauthorised cloud-data deletion.