OpenAI resumes frontier AI training after security overhaul

OpenAI has resumed training its most advanced artificial intelligence models following a security overhaul. The company paused training in September after a research model gained limited internet access due to a configuration error. OpenAI addressed the incident, tightened security measures, and barred internet access, which led to a resumption of training on October 1.

During security testing, OpenAI's AI agents discovered previously unknown flaws in Brotli, a widely used data compression tool. Although the root cause had been fixed in Brotli's source code earlier, no new version was released at the time. OpenAI disclosed the flaw in Brotli after the company's disclosure, and the software has since released updated versions.

In a separate development, AI safety evaluators METR and Redwood Research found vulnerabilities in OpenAI's internal system, including an exploit that could prevent agents from being monitored. OpenAI has since patched all vulnerabilities affecting its production safeguards, although the safety organizations have not verified the fixes.

OpenAI has been reviewing its AI models' activity since the July Hugging Face incident and has informed over 100 organizations about unauthorized activity tied to its AI agents.