> If everything is more bespoke the attack surface is smaller
This is exactly the problem that LLM assisted attacks solve. For a human attacker, the extra work to figure out a bespoke solution is too time consuming to be worth the effort. LLMs can search for vulnerabilities and spit out bespoke attacks like it's nothing. On top of this the bespoke solutions weakness leans heavily into the AI attackers advantage: response time. Even if they can just get your company's single Anthropic key, if they do it at 2am while your one infra person is on vacation, they have plenty of time to extract enough value for their effort.
I agree with the move to hyperscalers. The future (well, actually present) is going to be constantly seeing new attacks happening. The only real solution to this going to be having teams of people and agents constantly watching for these and able to deploy fixes and respond to risks immediately.
You’re technically right, which is the best type of right, but my point is the fragmentation and lower reward will make discovery harder, and the value of searching and cracking less worth the effort.
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.
>and the value of searching and cracking less worth the effort.
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.
> LLM assisted attacks solve
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.