OpenAI describes how it contains frontier model training and evaluation, including offline environments and high-compute red-teaming
On the same day Anthropic and the UK AISI reviewed an evaluation that went out of bounds online, OpenAI wrote about how it keeps its own training and evaluation contained.
// Key points
- Published on October 23 under OpenAI's Security category, the post explains how OpenAI strengthened containment for frontier model training and evaluation.
- The two measures named in OpenAI's summary are offline environments and high-compute red-teaming.
Builder's takeRead together with the AISI review on 10-24, two frontier labs made containing the model during testing a public practice in the same week. We don't train frontier models, but the lesson carries: run agent evals and automated regression offline by default against fake external services, and allowlist any network access explicitly.
// Background · from #OpenAI
Full timeline →- Oct 24 OpenAI shut down 12 legacy model snapshots including GPT-4, GPT-4 Turbo, o1 and o3-mini, plus 5 fine-tuning bases, on October 23
- Oct 23 OpenAI's Decisions API is generally available on gpt-6-luna, with up to 400 questions per request and prompt caching for shared input
- Oct 22 OpenAI's frontier models and Codex workflows are coming to Google Cloud through Model Garden and Google Cloud Marketplace