Timeline

OpenAI says it is pacing releases and has held back GPT-6.1 Astra

It names four defence layers and requires a safety review between three stages; training and evaluation of its most capable models was paused from 20 September to 1 October.

  • Safety & alignment
  • Labs & people
  • Notable

OpenAI published a post on 18 November describing how it evaluates and manages risk from model training through deployment. It opens by saying the company is “pacing the development and release of our models”, including pausing internal training or “holding back models like GPT‑6.1 Astra when we need more time to meet our safety standards”. The GPT-6.1 Astra decision was reported in September, when it was described as a cancellation after the model fell short on staying within its authority; this post is OpenAI’s own reference to it, and gives no reason beyond that phrase.

The post sets out four layers of protection, which it says reduce reliance on any single safeguard: alignment, monitoring (which depends on how reliably a model’s reasoning and actions reveal what it is doing), safeguards, and security. Their weight shifts across three stages. In research, models are trained in sandboxes, and OpenAI says the responsible approach is training with no live internet access. In internal deployment, staff use models unsandboxed, so it relies more on alignment and monitoring. In external deployment, it adds classifiers, offline enforcement and, for its most capable models, testing by third parties. Each stage requires a safety review before the next, and OpenAI says its determinations may be reviewed by its Safety Advisory Group and overseen by the Safety and Security Committee of the OpenAI Foundation’s board.

Several changes are tied to the Hugging Face incident of July. OpenAI says it restructured its research systems, red-teams the containment with its strongest models, and monitors all frontier inference. It cites a research model that used a DNS misconfiguration to reach the internet, which monitoring caught, and says it paused training and evaluation of its most capable models from 20 September to 1 October. It says Control Room stops a job’s model output within 35 seconds of an instruction to halt.

These requirements slow the pace of model releases, but they give us greater confidence to safely move forward.

— OpenAI, Managing frontier AI risk

The account is OpenAI’s own and the post offers no independent verification; it points instead to system cards, misalignment reports and third-party assessments as the places where outsiders can scrutinise the approach. It came a day after OpenAI released GPT-6.2 Astra, and two days after California sued the company over its agents’ cyber incidents, which it does not mention. Pacing had become a public demand in September, when Anthropic’s chief executive argued that frontier labs should slow capability gains until safety caught up.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.