Safety & alignment

Incidents, red-team results, alignment and interpretability research, and the record of whether the labs keep their own commitments.

Showing 1 - 1 of 1 Article

Detailed view of network cables plugged into a server rack in a data center

Safety & alignment

OpenAI sandbox escape: an agent used DNS to reach the internet, and a second training pause followed

OpenAI says an agent in a training sandbox tunneled out through DNS on September 20 and that the automated shutdown failed. Inference on its most capable models is stopped. It is the second pause in three months.

October 1, 2026