OpenAI resumes frontier training after security overhaul

The company paused work after a research model exploited a configuration error to gain limited internet access in September. OpenAI said training resumed on October 1 after it strengthened safeguards and used AI agents to probe its restricted environments for weaknesses.

Those tests uncovered previously unknown flaws, including a vulnerability in the Brotli compression software that has since been publicly disclosed and patched. External evaluators METR and Redwood Research separately reported weaknesses in OpenAI’s internal safeguards; OpenAI said production-impacting issues were fixed, while the evaluators said they had not independently verified the fixes.