I understand your concern. The way we designed the deployment tool was under the assumption people would be using a freshly-deployed cloud single-use server for it (as we assume they have no technical knowledge).
I'm not sure if a container is foolproof. There have been multiple CVEs in the past allowing processes to escape containers through kernel vulnerabilities. Although, I'm happy to put containers on our to-do list if this will help.
As for what the proper solution should be for advanced users, I personally am not sure. I'd need to research that further. We do try to provide things such as reproducible builds, which means if you build the code yourself using our reproducible build script, they'll match byte-for-byte against our released artifacts. This at least guarantees that it was built from our repository's code, although it does not guarantee the code itself is safe.
I think something that will help here is our planned third-party security audit, which hopefully will be sometime this summer.
The only thing AI is used on in this project is strictly for user interface work (our website, the front-end for the mobile app, the front-end for the deploy tool). We carefully vet anything like that.
I think you may have misinterpreted my "automations, systems, and AI" (you put it in bold), that is intended to show my experience in machine learning (example: I spent 4 months in a lab helping improve the accuracy of wearable ECG abnormality detection). I do not rely on LLMs.