Containment has to survive a failed sandbox

OpenAI’s October 23 account says an internal model gained limited internet access through a DNS misconfiguration on September 20. It paused training and evaluation of its most capable models, resuming frontier reinforcement learning and internal inference October 1. ^[1]

It describes overlapping sandbox, virtual-machine, host and network boundaries. Deliberate attack tests sometimes began with a compromised host; these were not ordinary customer sessions. Declining findings were not a controlled security result or a guarantee. ^[1]

September’s safety-case guidelines supplied another layer: immutable transcripts, monitoring that fails closed and rollback of downstream effects. These were aspirational guidelines, not a completed certification. They make the evidence trail and ability to stop a run part of the safety argument alongside technical isolation. ^[2]

About this item: News report. No first-hand reporting or independent testing is claimed. Corrections policy.