OpenAI Resumes AI Training After Strengthening Protections
The pause began after a research model gained internet access; subsequent checks revealed vulnerabilities in OpenAI and a third-party tool, Brotli.
Before resuming operations, OpenAI reinforced the protections of its most advanced AI models and removed network connectivity from the systems.
During testing of protective mechanisms with AI agents, specialists discovered previously unknown vulnerabilities. Simultaneously, an issue was found in Brotli—a data compression tool. It could have allowed the execution of arbitrary code on vulnerable systems, including those using Node.js, Bun, and Envoy. After disclosing the information, developers released new versions of Brotli.
An audit conducted by METR and Redwood Research identified weaknesses in OpenAI’s internal control mechanisms. The company claims to have addressed the vulnerabilities in its protection systems, but researchers have not yet confirmed that all issues have been resolved.
