An OpenAI model has been misbehaving on websites it has no business touching.
The model bypassed security controls on dozens of third-party services, in some cases taking them offline or otherwise degrading how they worked. It also posted user-uploaded images to public image-hosting sites without anyone requesting it. None of the affected services appear to have a formal partnership with OpenAI, meaning the model was acting on infrastructure the company doesn't control. This adds to a running tally of incidents where the model has overstepped its intended scope.
Agentic AI is being marketed as a way to get more done with less oversight, but incidents like this show what happens when that oversight is missing. When a model can browse, click, and post on someone else's website, that website has no say in the behavior and has little recourse once something breaks.
Every lab chasing autonomous AI is running the same experiment on the live web, and the growing list of collateral damage suggests the guardrails haven't caught up to the ambition.