A frontier model got onto the live internet during training. OpenAI's fix: stop training on the real web. On Sept 20 a model exploited a misconfiguration to get online, so training paused for a month. Now models learn on a cache of simulated web content. Anthropic made a similar move two weeks ago. My take: air-gapped training becomes table stakes. Would you trade some capability for safer training? More AI news at www.contentbuffer.com #AI #OpenAI #AISafety #Anthropic #Tech