Comment by nilsb

That's not the point of why those environments are sandboxed. They're restricted so that during training the LLM doesn't "cheat", for example by looking up answers on Google (simplified example). You wouldn't want an LLM that responds to "fix this bug in my repo" with replacing your code with something it found on the web.