The prominent game right now is about the classes of attacks that AI can recognize both to exploit and to remedy.

Implicitly, that creates a new highly valuable class of vulnerabilities: the ones that AI for whatever reason is unable to readily see. Understanding those two groups will be paramount and will lead to better AI.

Don't run on open source software - doesn't matter how obscure it is, the LLMs will know about it.

Don't run on software that other people use - if there is more than one copy of the software accessible, it'll be broken elsewhere and you'll be a victim of drive-by hacking.

Do - run your own software, written by you, which nobody else uses, that is closed source and completely undocumented. Then you might have some chance of avoiding being hacked.

I'm using AI to decompile old arcade games.

I am finding the craziest anti-tampering stuff. Three different routines - each watches the code of the others - and if anything changes - tank the machine by jumping into graphics RAM.

When I first started seeing this stuff I assumed my decompiler was broken. Nope. Just insane code.

You can absolutely make an agent lose its mind thus making it spend more tokens, to an extent where it might not be viable, but in all case, slowing down or making something more expensive is a direct way to increase security, so wasting either token or human time both is useful.

SBO+Real security > Security.