It seems almost impossible. I spent the weekend comparing nanoclaw to openclaw - nanoclaw is a slightly more secure version - containerized filesystem basically - and very popular.

It's a) harder to setup, b) less functional out of the box, c) has almost exactly the same security risk surface -- either you hook it up to your email, comms, documents and give it API tokens, or you don't. If you do -- well, at least it can't delete your hard drive without turning full evil and looking for red pill type exploits that break the container -- but, it still has the same other security dynamics.

Anyway, employing a very suspicious watcher that's hooked to the shell and API calls is probably the way forward. Can that thing be reasoned with / tricked?

The same way as delivering a "truly secure human". Which is of course impossible, that's why spies, double agents and even triple agents exist and have succeeded to do their job. And that's despite an enormous number of guardrails meant to prevent exactly that.

And simply "secure enough" doesn't help much either, because whereas a single human spy can only do so much damage, if an LLM is given access to everything in one way or another - which is the whole concept - then the potential damage is boundless.

That's easy. We just keep pumping these things and remind everyone that there's no real consequences (at least to the people who actually matter) and what was previously agreed as super important and critical will eventually turn out to no longer be super important or critical. Lethal trifecta solved. Who cares if your agent is forwarding private and confidential emails to random people, if everyone else is doing it too. Syndrome from the Incredibles movie won, and we helped make it happen. In fact, we made sure of it.

There are secure alternatives but they are not making millions of dollars.