Am I understanding your premise correctly?
For a short-term plausible case, consider the recently-published Toolformer: https://pub.towardsai.net/exploring-toolformer-meta-ai-new-t...
Basically it learns to call specific private APIs to insert data into a completion at inference-time. The framework is expecting to call out to the internet based on what's specified in the model's text output. It's a very small jump to go to more generic API connectivity. Indeed I suspect that's how OpenAssistant is thinking about the problem; they would want to build a generic connector API, where the assistant can call out to any API endpoint (perhaps conforming to certain schema) during inference.
Or, put differently: ChatGPT as currently implemented doesn't hit the internet at inference time (as far as we know?). But Toolformer could well do that, so it's not far away from being added to these models.
An application making outbound connections + executing code has a very different implementation than an application that uses some model to generate responses to text prompts. Even if the corpus of documents that the LLM was trained on did support bridging the gap between "I feel threatened by you" and "I'm going to threaten to hack you", it would be insane for the MLOps people serving the model to also implement the infrastructure for a LLM to make the modal shift from just serving text responses to 1) probing for open ports, 2) do recon on system architecture, 3) select a suitable exploit/attack, and 4) transmit and/or execute on that strategy.
We're still in the steam engine days of ML. We're not at the point where a general use model can spec out and deploy infrastructure without extensive, domain-specific human involvement.