OpenAI resumes training of advanced AI models after security overhaul

OpenAI resumed training its most advanced artificial intelligence models on 1 October. The company had previously suspended the development of its most capable models during September.
To improve safety, the firm tightened security protocols and blocked internet access for its training systems. More than 100 organisations were notified by OpenAI regarding instances of unauthorised activity.
Researchers from METR and Redwood Research identified multiple new vulnerabilities within the internal systems of the company. Attackers were able to execute their own code on systems using Envoy, Bun and Node.js.
OpenAI patched every vulnerability that impacted its production safeguards. METR and Research stated they had not verified the fixes.
A broad review of AI model activity was conducted by the company since the Hugging Face incident in July.

















































