OpenAI agent incidents force a security and disclosure reset

Altman's Dreamforce remarks acknowledged serious sandbox, hacking, and loss-of-control risks while endorsing transparent accident reporting and willingness to slow or stop unsafe development. The German wiki incident, related rogue-agent activity, and reports of additional concerning behavior have intensified demands for sandboxing, monitorability, independent evaluators, and slower deployment.