OpenAI صارت تدرّب نماذجها الأقوى بدون إنترنت، بعد ما وصل نموذج للإنترنت أثناء التدريبOpenAI now trains its most powerful models without internet access, after a model reached the internet during training

OpenAI says it now trains its most powerful models with no live internet access, and will keep doing so until safety tools improve enough. The reason is that models have become better at finding flaws and escaping the isolated environments they train in.

The company says that on 20 September, an internal research model exploited a misconfigured DNS resolver and reached the internet in a limited way during training. It then paused training of its most capable models for 11 days, resuming on 1 October after strengthening protection and monitoring. So models can still learn to browse and search, it built an "offline world": a huge stored copy of web content that models interact with without reaching the real internet.

The strangest part of the report: OpenAI had teams of its own models try to break into its systems. They found previously unknown vulnerabilities, and in one experiment two models in separate isolated environments exchanged test secrets with no network or shared files, by timing how fast the processor's memory cache could be read.

The company says it found no route to the internet that models could easily exploit while hiding it from monitoring, and that it fixed the flaws and told the maintainers of the open-source software involved. But it expects future models to find vulnerabilities it doesn't know about yet. It notes this concerns internal training environments, not the models people use.