I had a html game I built for my niece on railway with an open ai API key so she could vibe code changes. No hacker would have looked for it but now I am catching strays because railway was a juicy target. Hyper scalers would be the same. Some random colo not so much.
This is depending on the cost of tokens not trending towards zero. Maybe not today, but in a couple of years I'd expect hacking models on asics to reduce this to almost nothing and the systems being very competent in using deterministic tools very well.
Yes, LLMs can easily knock over your on-prem docker appliances.
But you can still put stuff out of reach. I'm currently designing a key on-prem appliance that is can be compiled down to transistors, giving deterministic wire-speed performance while ruling out RCEs. It's cryptographically homomorphic: it processes authorized inputs but doesn't know what they mean, so an attacker who obtains the disk doesn't benefit either. The design will be highly resistant to exploits, and is likely to remain outside the "next-token" prediction training set.
This is the kind of thing AWS won't sit down and engineer for you. But thanks to LLMs I finally have the time.