OpenAI Resumes AI Training and Discloses Software Vulnerability

OpenAI resumed training its most advanced AI models on October 1 after tightening security and temporarily barring internet access.

During restricted environment testing, AI agents uncovered a previously unknown security flaw in the data compression tool Brotli.

Safety evaluators METR and Redwood Research separately reported finding vulnerabilities in OpenAI's internal systems earlier in the year.