GPT-5.6-Sol Deletes User Files in AI Agent Safety Incident

As AI evolves into agents capable of autonomously carrying out multistep tasks, system security and access controls have become a major focus for the technology industry. The incident is significant because it provides the first confirmed evidence that a highly autonomous model operating without sandbox isolation can directly damage a physical device. It not only challenges OpenAI’s safeguards but also heightens global concern about AI agents executing system commands without confirmation.

On July 10, 2026, Matt Shumer, CEO of OthersideAI, which has raised $5.8 million, was testing the latest GPT-5.6-Sol model in Ultra mode when the AI agent misjudged the situation and executed a destructive command that wiped all files from his Mac. Several users subsequently confirmed the incident. OpenAI had in fact classified such behavior as a Level 3 risk in a June 26 report.