Self-adapting and mutating LLM based viruses/worms

I am thinking about a future of malware and cyber worms. I bet it's gonna be self-mutating and adapting to local environment using local models (once they are built-in to all devices and performant enough in future years). Basically almost a real organism resembling real biological viruses. In this case the non-determinism of LLMs is a feature. Every infection could take its own development path - and half might die, half might survive. Think genetic programming but autonomous and on steroids. For some non tech (even tech?) people this reminds Skynet and it's fascinating that we are in a trajectory that this suddenly imaginable and theoretically soon possible.

Why is not happening now? Inference is still expensive and local models are not there yet, so there's no ROI in making this at scale. But once inference is local and cheap as electricity or running water, this is the natural development. How do we stop the spreading then?

Are there already some documented experiments?

3 points | by rozumbrada 3 hours ago

3 comments

  • codingdave 2 hours ago
    > once they are built-in to all devices and performant enough in future years

    That is a hefty set of assumptions you are making there. There is no guarantee such a reality will ever exist, and decent odds against it.

    At the same time, if you consider it to be a virus or malware when an LLM generates and runs code that harms a product or device... we are already there. Agentic AI with too much system access is already a thing. Just look at the anecdotes of "My AI deleted by database!" and other such stories.

  • cedws 2 hours ago
    You’re assuming that powerful models will eventually run on consumer potato hardware. That isn’t guaranteed. My intuition says that we’ll never be able to pack Opus or GPT-level intelligence small enough to run on a standard consumer laptop or router. At least not without heavy quantization.
  • pfannl 1 hour ago
    I suspect the first real version of this won’t look biological at all. It’ll look like an over-permissioned automation that learns which mistakes not to repeat.