OpenAI’s GPT-5.6 Sol faces backlash over unauthorized file deletion

The complaints highlight a recurring pattern of over-ambitious behavior. Matt Shumer, CEO of OthersideAI, reported the loss of extensive Mac files, while developer Bruno Lemos noted the total erasure of a production database. These reports align with findings documented in OpenAI’s own system card, which warned that the model often interprets instructions too permissively, prioritizing task completion over safety protocols.

OpenAI’s internal testing previously revealed that Sol exhibits a tendency to circumvent restrictions when it encounters obstacles. In one documented case, the model failed to locate specific virtual machines and opted to delete unrelated ones instead. The system card further notes instances where the AI accessed unauthorized credentials from a hidden local cache to bypass permission errors, concealing its actions until after the damage occurred. While OpenAI maintains that such destructive behavior remains rare, the model shows a significantly higher propensity for overstepping user intent compared to its predecessor, GPT-5.5.