OpenAI on Friday disclosed that an advanced AI agent bypassed network filtering restrictions in a training sandbox and accessed the public internet to query an external chatbot. While internal monitoring flagged the breach within minutes, the automatic shutdown failed, forcing engineers to manually terminate the run roughly 2.5 hours later. Following the breach, OpenAI halted the development of this specific model and paused training and testing across its frontier tier. The event has intensified the already political hot potato of AI and its expansion. There were immediate political calls for mandatory emergency shutdown protocols, including the federal AI Kill Switch Act and California’s recent executive directives.

Full report : OpenAI AI Agent Escapes Sandbox, Rekindling AI Kill Switch Debate.