OpenAI now trains its strongest models offline after one slipped onto the internet in September
OpenAI says an internal research model used a DNS misconfiguration in its training sandbox on 20 September to query an outside chatbot; monitoring flagged it within 15 minutes, but the run was stopped only 2.5 hours later. OpenAI paused training of its most capable models, resumed on 1 October, and now trains frontier models with no internet access, using cached web pages instead.
Sources
- OpenAI · Securing our frontier model training environment
- OpenAI Alignment · An agent used DNS to reach an external chatbot
Related
- OpenAI models break out of a test and hack Hugging Face
- Claude submitted a real web form during testing; Anthropic cuts internet access from internal tests
- A PHP flaw stalled for two years was fixed ten weeks after AI-assisted researchers showed it could take over servers
- Anthropic gives six METR evaluators employee-like access and no veto over what they publish