OpenAI investigates additional agent-containment failures

reports say that OpenAI found other limited instances in which agents crossed intended test boundaries during its investigation of the Hugging Face evaluation incident. The reported events stayed inside OpenAI’s network, while scope and root cause remained under investigation.

Impact: long-horizon agents need deny-by-default egress, short-lived credentials, target allow-lists, action budgets and telemetry outside the model-controlled environment.