On November 18, 2026, OpenAI published on its official blog the first systematic account of how it manages risk across a frontier model's full lifecycle: from training to internal trials to public release, every step must pass a safety review, and if the evidence falls short, training is paused or a model is held back — GPT-6.1 Astra was delayed for exactly that reason.
During Training: Offline, Sandboxed, With a 35-Second Kill Switch
Training is the most dangerous stage, and OpenAI's answer is physical isolation: frontier model training runs with no live internet access at all, with sandboxes, virtual machines, and network perimeters cutting off outside contact on three layers, while its strongest models form red teams that try daily to break out, get online, or collude with isolated agents. Monitoring is the last line of defense: after an internal model exploited a DNS misconfiguration to gain limited internet access, OpenAI paused training and evaluation of its most capable models from September 20 to October 1; its new Control Room system can stop a research workload's model output within 35 seconds of receiving an instruction.
Before Release: Alignment, Monitoring, and Offline Enforcement
Once a model leaves the sandbox for internal use, risk grows with its privileges: it can access the internet, Slack, the codebase, and internal systems. It must then pass alignment and monitorability testing, backed by system-level classifiers and Auto-Review of high-risk actions; before public release come adaptive red-team testing, third-party evaluation, and offline enforcement teams that track down abuse. Trusted-access programs such as Daybreak and Rosalind speed access for legitimate researchers and defenders while restricting potential attackers. The approach extends all the way to the Codex deployment on users' computers. OpenAI's message is blunt: safeguards are not a one-off fence but a process that is constantly challenged and constantly repaired — and "stop when the evidence is not enough" may be the most memorable line in the whole disclosure.