On 2024/06/12 13:37, Otto Moerbeek wrote:
> On Wed, Jun 12, 2024 at 04:28:05AM -0300, Alfredo Ortega wrote:
> 
> > The 10000 patches number is just for the IPV4/IPV6 stack. I also don't
> > think you should review or integrate them, because in a couple months
> > when more advanced LLMs are made available I can regenerate all the
> > patches in less than a morning with much better quality. And again
> > every time a new LLM is released.
> > 
> > That's why I think of the patches as a post-processing step. I.E. you
> > keep the regular process of development, and I or other people can
> > refactor and release secure versions of the kernel/userland.
> 
> You have *not* demonstrated your patches will produce a more secure
> version of the code. That's just a big assumption you made with zero
> evidence.

And even if you do, that will need re-doing "when more advanced LLMs are
made available" and the patches are updated "in less than a morning".

> > It's great that you want to keep the development process human, but my
> > opinion is that if you have AI adversaries (like we have now), you
> > need AI protections.
> 
> Again, you assume AI will provide protection.

Threat model needs to include "deliberately poisoning LLM source
material so that it deliberately introduces vulnerabilities". (And of
course it would be super easy to sneak something looking like some of
the changes involved with the xz backdoor amongst a bunch of machine-
generated modifications...)