OpenAI resumes advanced AI training after security overhaul

OpenAI resumed training its most advanced AI models on October 1 after a security pause prompted by a research model gaining unauthorized internet access in September.

During defense testing, AI agents uncovered a previously unknown security flaw in Brotli, a widely used data compression tool.

Safety evaluators METR and Redwood Research separately disclosed finding internal vulnerabilities, though OpenAI stated it has since patched production safeguards.