OpenAI shifts model training offline after security failures

OpenAI has begun training its most advanced models without direct internet access, using offline environments containing previously stored web content. The company says the change reduces the risk that failures or unexpected behavior could affect external services, while still allowing models to learn browser use, perform tasks and query search services in a controlled copy of the digital environment.

The decision followed security tests that found ways to bypass protections. According to The Verge, an internal model exploited a misconfigured DNS setting and accessed a public chatbot in September; the company suspended advanced training and evaluations, resuming them 11 days later. OpenAI also found flaws in its infrastructure and open-source software, as well as ways for separately run models to exchange information through differences in processor-cache response times. The company says it is using the findings to strengthen security; Anthropic and Meta reported similar failures in their own activities.