The AI tries to follow the style of the existing checks in the code,
but I can easily tell it to panic in case of a security fail.
And I do not plan to submit this particular batch of checks, and they
will become obsolete in about a month when the next gen AIs are made
public.
Most of the checks of this refactor are being done with GPT-4, that is
not even the best current coding AI. And the mechanism of patching is
crude, at best. Yet, it works.
I may be wrong, but I believe by this time next year the AI will be so
good that I doubt I will even need human reviewers.

El mar, 11 jun 2024 a las 9:54, Stuart Henderson
(<stu@spacehopper.org>) escribió:
>
> On 2024/06/11 09:28, Alfredo Ortega wrote:
> > I added 10000+ checks so far, in about 4 or 5 hs. Final count will
> > likely be close to a million.
> > It's true that many are useless, perhaps up to 50% of them.  Most
> > stack protections put into place by the compiler are also useless.
> > But the question is, how many are not useless? and how many checks
> > humans missed, but the AI correctly put in place?
>
> Seems that many of the checks are adding return/continue when things
> don't match conditions which aren't handled in the code. But who is to
> say that's a safe thing to do in any given case? It might often be
> better to let the kernel crash so the problems are more obvious.
>